Category Archives: Ai

Along with unsupervised machine learning and supervised learning, another common form of AI creation is reinforcement learning. Beyond regular reinforcement learning, deep reinforcement learning can lead to astonishingly impressive results, thanks to the fact that it combines the best aspects of both deep learning and reinforcement learning. Lets take a look at precisely how deep reinforcement learning operates. Note that this article wont delve too deeply into the formulas used in deep reinforcement learning, rather it aims to give the reader a high level intution for how the process works.

Before we dive into deep reinforcement learning, it might be a good idea to refresh ourselves on how regular reinforcement learning works. In reinforcement learning, goal-oriented algorithms are designed through a process of trial and error, optimizing for the action that leads to the best result/the action that gains the most reward. When reinforcement learning algorithms are trained, they are given rewards or punishments that influence which actions they will take in the future. Algorithms try to find a set of actions that will provide the system with the most reward, balancing both immediate and future rewards.

Reinforcement learning algorithms are very powerful because they can be applied to almost any task, being able to flexibly and dynamically learn from an environment and discover possible actions.

Photo: Megajuice via Wikimedia Commons, CC 1.0 (https://commons.wikimedia.org/wiki/File:Reinforcement_learning_diagram.svg)

When it comes to deep reinforcement learning, the environment is typically represented with images. An image is a capture of the environment at a particular point in time. The agent must analyze the images and extract relevant information from them, using the information to inform which action they should take. Deep reinforcement learning is typically carried out with one of two different techniques: value-based learning and policy-based learning.

Value-based learning techniques make use of algorithms and architectures like convolutional neural networks and Deep-Q-Networks. These algorithms operate by converting the image to greyscale and cropping out unnecessary parts of the image. Afterward, the image undergoes various convolutions and pooling operations, extracting the most relevant portions of the image. The important parts of the image are then used to calculate the Q-value for the different actions the agent can take. Q-values are used to determine the best course of action for the agent. After the initial Q-values are calculated, backpropagation is carried out in order that the most accurate Q-values can be determined.

Policy-based methods are used when the number of possible actions that the agent can take is extremely high, which is typically the case in real-world scenarios. Situations like these require a different approach because calculating the Q-values for all the individual actions isnt pragmatic. Policy-based approaches operate without calculating function values for individual actions. Instead, they adopt policies by learning the policy directly, often through techniques called Policy Gradients.

Policy gradients operate by receiving a state and calculating probabilities for actions based on the agents prior experiences. The most probable action is then selected. This process is repeated until the end of the evaluation period and the rewards are given to the agent. After the rewards have been dealt with the agent, the networks parameters are updated with backpropagation.

Because Q-Learning is such a large part of the deep reinforcement learning process, lets take some time to really understand how the Q-learning system works.

The Markov Decision Process

A markov decision process. Photo: waldoalvarez via Pixabay, Pixbay License (https://commons.wikimedia.org/wiki/File:Markov_Decision_Process.svg)

In order for an AI agent to carry out a series of tasks and reach a goal, the agent must be able to deal with a sequence of states and events. The agent will begin at one state and it must take a series of actions to reach an end state, and there can be a massive number of states existing between the beginning and end states. Storing information regarding every state is impractical or impossible, so the system must find a way to preserve just the most relevant state information. This is accomplished through the use of a Markov Decision Process, which preserves just the information regarding the current state and the previous state. Every state follows a Markov property, which tracks how the agent change from the previous state to the current state.

Deep Q-Learning

Once the model has access to information about the states of the learning environment, Q-values can be calculated. The Q-values are the total reward given to the agent at the end of a sequence of actions.

The Q-values are calculated with a series of rewards. There is an immediate reward, calculated at the current state and depending on the current action. The Q-value for the subsequent state is also calculated, along with the Q-value for the state after that, and so on until all the Q-values for the different states have been calculated. There is also a Gamma parameter that is used to control how much weight future rewards have on the agents actions. Policies are typically calculated by randomly initializing Q-values and letting the model converge toward the optimal Q-values over the course of training.

Deep Q-Networks

One of the fundamental problems involving the use of Q-learning for reinforcement learning is that the amount of memory required to store data rapidly expands as the number of states increases. Deep Q Networks solve this problem by combining neural network models with Q-values, enabling an agent to learn from experience and make reasonable guesses about the best actions to take. With deep Q-learning, the Q-value functions are estimated with neural networks. The neural network takes the state in as the input data, and the network outputs Q-value for all the different possible actions the agent might take.

Deep Q-learning is accomplished by storing all the past experiences in memory, calculating maximum outputs for the Q-network, and then using a loss function to calculate the difference between current values and the theoretical highest possible values.

Deep Reinforcement Learning vs Deep Learning

One important difference between deep reinforcement learning and regular deep learning is that in the case of the former the inputs are constantly changing, which isnt the case in traditional deep learning. How can the learning model account for inputs and outputs that are constantly shifting?

Essentially, to account for the divergence between predicted values and target values, two neural networks can be used instead of one. One network estimates the target values, while the other network is responsible for the predictions. The parameters of the target network are updated as the model learns, after a chosen number of training iterations have passed. The outputs of the respective networks are then joined together to determine the difference.

Policy-based learning approaches operate differently than Q-value based approaches. While Q-value approaches create a value function that predicts rewards for states and actions, policy-based methods determine a policy that will map states to actions. In other words, the policy function that selects for actions is directly optimized without regard to the value function.

Policy Gradients

A policy for deep reinforcement learning falls into one of two categories: stochastic or deterministic. A deterministic policy is one where states are mapped to actions, meaning that when the policy is given information about a state an action is returned. Meanwhile, stochastic policies return a probability distribution for actions instead of a single, discrete action.

Deterministic policies are used when there is no uncertainty about the outcomes of the actions that can be taken. In other words, when the environment itself is deterministic. In contrast, stochastic policy outputs are appropriate for environments where the outcome of actions is uncertain. Typically, reinforcement learning scenarios involve some degree of uncertainty so stochastic policies are used.

Policy gradient approaches have a few advantages over Q-learning approaches, as well as some disadvantages. In terms of advantages, policy-based methods converge on optimal parameters quicker and more reliably. The policy gradient can just be followed until the best parameters are determined, whereas with value-based methods small changes in estimated action values can lead to large changes in actions and their associated parameters.

Policy gradients work better for high dimensional action spaces as well. When there is an extremely high number of possible actions to take, deep Q-learning becomes impractical because it must assign a score to every possible action for all time steps, which may be impossible computationally. However, with policy-based methods, the parameters are adjusted over time and the number of possible best parameters quickly shrinks as the model converges.

Policy gradients are also capable of implementing stochastic policies, unlike value-based policies. Because stochastic policies produce a probability distribution, an exploration/exploitation trade-off does not need to be implemented.

In terms of disadvantages, the main disadvantage of policy gradients is that they can get stuck while searching for optimal parameters, focusing only on a narrow, local set of optimum values instead of the global optimum values.

Policy Score Function

The policies used to optimize a models performance aim to maximize a score function J(). If J() is a measure of how good our policy is for achieving the desired goal, we can find the values of that gives us the best policy. First, we need to calculate an expected policy reward. We estimate the policy reward so we have an objective, something to optimize towards. The Policy Score Function is how we calculate the expected policy reward, and there are different Policy Score Functions that are commonly used, such as: start values for episodic environments, the average value for continuous environments, and the average reward per time step.

Policy Gradient Ascent

Gradient ascent aims to move the parameters until they are at the place where the score is highest. Photo: Public Domain (https://commons.wikimedia.org/wiki/File:Gradient_ascent_(surface).png)

After the desired Policy Score Function is used, and an expected policy reward calculated, we can find a value for the parameter which maximizes the score function. In order to maximize the score function J(), a technique called gradient ascent is used. Gradient ascent is similar in concept to gradient descent in deep learning, but we are optimizing for the steepest increase instead of decrease. This is because our score is not error, like in many deep learning problems. Our score is something we want to maximize. An expression called the Policy Gradient Theorem is used to estimate the gradient with respect to policy .

In summary, deep reinforcement learning combines aspects of reinforcement learning and deep neural networks. Deep reinforcement learning is done with two different techniques: Deep Q-learning and policy gradients.

Deep Q-learning methods aim to predict which rewards will follow certain actions taken in a given state, while policy gradient approaches aim to optimize the action space, predicting the actions themselves. Policy-based approaches to deep reinforcement learning are either deterministic or stochastic in nature. Deterministic policies map states directly to actions while stochastic policies produce probability distributions for actions.

See the article here:

AI Used to Monitor Health of Coral Reefs and Detect Ocean Trash Pollution - Unite.AI

LONDON--(BUSINESS WIRE)--The artificial intelligence (AI) market in retail sector is expected to grow by USD 14.05 billion during 2019-2023. The report also provides the market impact and new opportunities created due to the COVID-19 pandemic. The impact can be expected to be significant in the first quarter but gradually lessen in subsequent quarters with a limited impact on the full-year economic growth, according to the latest market research report by Technavio. Request a free sample report

Companies operating in the retail sector are increasingly adopting AI solutions to improve efficiency and productivity of operations through real-time problem-solving. For instance, the integration of AI with inventory management helps retailers to effectively plan their inventories with respect to demand. AI also helps retailers to identify gaps in their online product offerings and deliver a personalized experience to their customers. Many such benefits offered by the integration of AI are crucial in driving the growth of the market.

To learn more about the global trends impacting the future of market research, download a free sample: https://www.technavio.com/talk-to-us?report=IRTNTR31763

As per Technavio, the increased applications in e-commerce will have a positive impact on the market and contribute to its growth significantly over the forecast period. This research report also analyzes other significant trends and market drivers that will influence market growth over 2019-2023.

Artificial Intelligence (AI) Market in Retail Sector: Increased Applications in E-commerce

E-commerce companies are increasingly integrating AI in various applications to gain a competitive advantage in the market. The adoption of AI-powered tools helps them to analyze the catalog in real-time to serve customers with similar and relevant products. This improves both sales and customer satisfaction. E-commerce companies are also integrating AI with other areas such as planning and procurement, production, supply chain management, in-store operations, and marketing to improve overall efficiency. Therefore, the increasing application areas of AI in e-commerce is expected to boost the growth of the market during the forecast period.

Bridging offline and online experiences and the increased availability of cloud-based applications will further boost market growth during the forecast period, says a senior analyst at Technavio.

Register for a free trial today and gain instant access to 17,000+ market research reports

Technavio's SUBSCRIPTION platform

Artificial Intelligence (AI) Market in Retail Sector: Segmentation Analysis

This market research report segments the artificial intelligence (AI) market in retail sector by application (sales and marketing, in-store, planning, procurement, and production, and logistics management) and geographic landscape (North America, APAC, Europe, MEA, and South America).

The North America region led the artificial intelligence (AI) market in retail sector in 2018, followed by APAC, Europe, MEA, and South America respectively. During the forecast period, the North America region is expected to register the highest incremental growth due to factors such as early adoption of AI, rising investments in R&D and start-ups, and increasing investments in technologies.

Technavios sample reports are free of charge and contain multiple sections of the report, such as the market size and forecast, drivers, challenges, trends, and more. Request a free sample report

Some of the key topics covered in the report include:

Market Drivers

Market Challenges

Market Trends

Vendor Landscape

About Technavio

Technavio is a leading global technology research and advisory company. Their research and analysis focus on emerging market trends and provides actionable insights to help businesses identify market opportunities and develop effective strategies to optimize their market positions.

With over 500 specialized analysts, Technavios report library consists of more than 17,000 reports and counting, covering 800 technologies, spanning across 50 countries. Their client base consists of enterprises of all sizes, including more than 100 Fortune 500 companies. This growing client base relies on Technavios comprehensive coverage, extensive research, and actionable market insights to identify opportunities in existing and potential markets and assess their competitive positions within changing market scenarios.

More here:

Pre & Post COVID-19 Market Estimates-Artificial Intelligence (AI) Market in Retail Sector 2019-2023| Increased Efficiency of Operations to Boost...

Built Robotics executives are fond of saying that their autonomous system for construction equipment, like dozers and excavators, might be further along than many autonomous vehicles. In fact, CEO Noah Ready-Campbell insists youll see autonomous vehicles in controlled industrial environments like construction sites before you see level 5 driverless cars on public roads. That may be in part because autonomous construction equipment often operates on privately owned land, while public roads face increased regulatory scrutiny.

Theres a quote that Cold fusion is 20 years in the future and always will be,' Ready-Campbell told VentureBeat. I think theres a chance that that might be true for level 5 self-driving cars as well.

That might have seemed like an absurd thing to say back when autonomous driving first entered the collective imagination and companies established their intention to solve AIs grand autonomous vehicle challenge. But Waymo now takes billions from outside investors, and the delay of major initiatives like GMs Cruise and taxi service and Fords autonomous driving program call into question the progress automakers have made on autonomous vehicles.

One thing Ready-Campbell credits autonomous vehicle companies with is generating excitement around AI for use in environments beyond public roads, like on construction sites.

We were the beneficiaries of that when we did our series B last year, he said. I definitely think construction benefited from that.

From computer vision systems and drones to robots walking and roving through construction projects, Built Robotics and a smattering of other companies are working in unstructured industrial environments like mining, agriculture, and construction to make autonomous systems that can build, manage, and predict outcomes.

To take a closer look at innovation in the field, the challenges ahead, and what its going to take to create fully autonomous construction projects in the future, VentureBeat spoke with startups that are already automating parts of their construction work.

Built Robotics creates control systems for existing construction equipment and is heavily focused on digging, moving, and placing dirt. The company doesnt make its own heavy construction equipment; its solution is instead a box of tech mounted inside heavy equipment made by companies like Caterpillar, Komatsu, and Hyundai.

Built Robotics VP of strategy Gaurav Kikani told VentureBeat that the company started with autonomous skid steers the little dozers that scoop up and transport sand or gravel on construction sites. Today, Built Robotics has autonomous systems for bulldozers and 40-ton excavators.

We have a software platform that actuates the equipment that takes all the data being read by the sensors on the machine every second and then makes decisions and actuates the equipment accordingly, Kikani said.

Built Robotics focuses on earthmoving projects at remote job sites in California, Montana, Colorado, and Missouri far removed from human construction workers. Autonomous heavy equipment monitored by a human overseer tills the earth in preparation for later stages of construction, when human crews arrive to do things like build homes or begin wind or solar energy projects. In the future, the startup, which raised $33 million last fall, wants to help with more infrastructure projects.

Kikani and Built Robotics CEO Ready-Campbell say the company is currently focused on projects where theres a lot of dirt to move but not a lot of qualified operators of heavy machinery.

Calling to mind John Henry versus the machine, Kikani said human operators can go faster than a Built-controlled excavator, for example, but machine automation is meant to provide consistency and maintain a reliable pace to ensure projects finish on schedule.

Built Robotics combines lidar with cameras for perception and to recognize humans or potential obstacles. Geofencing keeps machinery from straying outside the footprint of a construction site. Excavators and dozers can work together, with dozers pushing material away or creating space for the excavator to be more productive.

The fleet coordination element here is going to be critical. In Built [Robotic]s early days, we really focused on standalone activities, where you have one piece of equipment just on its own taking care of the scope. But realistically, to get into the heart of construction, I think were going to start to coordinate with other types of equipment, Kikani said. So you might have excavators loading trucks [and] autonomous haulage routes where you have fleets of trucks that are all kind of tracking along the same route talking to each other, alerting each other to what they see along the route if conditions are changing.

I think the trickiest thing about construction is how dynamic the environment is, building technology that is pliable or versatile enough to account for those changing conditions and being able to update in real time to plan to accommodate for that. I think that is really going to be the key here, he said.

Equipment operated by systems from companies like Built Robotics will also need computer vision to recognize utility lines, human remains, or anomalies like archeological or historically important artifacts. Its not an everyday occurrence, but construction activity in any locale can unearth artifacts that lead to work stoppage.

Drones that can deploy automatically from a box are being developed for a variety of applications, from fire safety to security to power line inspection. Drones hovering above a construction site can track project progress and could eventually play a role in orchestrating the movement of people, robotic equipment, and heavy machinery.

In a nod to natural systems, San Francisco-based Sunflower Labs calls its drones bees, its motion and vibration sensors sunflowers, and the box its drones emerge from a hive.

Sensors around a protected property detect motion or vibrations and trigger the drones to leave their base station and record photos and video. Computer vision systems working with sensors on the ground guide the drone to look for Intruders or investigate other activity. Autonomous flight systems are fixed with sensors on all four sides to influence where the drone flies.

Sunflower Labs CEO Alex Pachikov said his companys initial focus is on the sale of drones-in-a-box for automated security at expensive private homes. The company is also seeing a growing interest from farmers of high-value crops, like marijuana.

Multiple Sunflower Labs drones can also coordinate to provide security for a collection of vacation homes, acting as a kind of automated neighborhood watch that responds to disturbances during the months of the year when the homes attract few visitors.

Stanley Black and Decker, one of the largest security equipment providers in the United States, became a strategic investor in Sunflower Labs in 2017 and then started exploring how drones can support construction project security and computer vision services. Pachikov said Sunflowers security is not intended to replace all other forms of security, but to add another layer.

The companys system of bees, hives, and sunflowers is an easy fit for construction sites, where theft and trespassing at odd hours can be an issue, but the tools can do a lot more than safeguard vacant sites.

When a Sunflower Labs drone buzzes above a construction site, it can deploy computer vision-enabled analytics tools for volumetric measurement to convert an image of a pile of gravel into a prediction of total on-site material.

Then tools from computer vision startups like Pics 4D, Stockpile Reports, and Drone Deploy can provide object detection, 3D renderings of properties for tracking construction progress, and other image analysis tools.

Companies like Delair take a combination of data from IoT sensors, drone footage, and stationary cameras from a construction project to create a 3D rendering that Delair calls a digital twin. The rendering is then used to track progress and identify anomalies like cracks or structural issues.

Major construction companies around the world are increasingly turning to technology to reduce construction project delays and accident costs. The 2019 KPMG global construction survey found that within the next five years, 60% of executives at major construction companies plan to use real-time models to predict risks and returns.

Indus.ai is one of a handful of companies making computer vision systems for tracking progress on construction sites.

We can observe and use a segmentation algorithm to basically know every pixel what material it is and therefore we know the pace of your concrete work, your rebar work, your form work and [can] start predicting whats happening, Indus.ai CEO Matt Man told VentureBeat in a phone interview.

He envisions robotic arms being used on construction sites to accomplish a range of tasks, like creating materials or assembling prefabricated parts. Digitization of data with sensors in construction environments will enable various machine learning applications, including robotics and the management of environments with a mix of working humans and machines.

For large projects, cameras can track the flow of trucks entering a site, the number of floors completed, and the overall pace of progress. Computer vision could also follow daily work product and help supervisors determine whether the work of individuals and teams follows procedure or best trade practices.

Imagine a particular robotic arm can start putting drywall up, then start putting tiles up, all with one single robotic arm. And thats where I see the future of robotics [] To be able to consolidate various trades together to simplify the process, Man said. There could be armies of robot-building things, but then there is an intelligent worker or supervisor who can manage five or 10 robotic arms at the same time.

Man thinks software for directing on-site activity will become more critical as contractors embrace robotics, and he sees a huge opportunity for computer vision to advance productivity and safety in industrial spaces.

Stanford University engineers have explored the use of drones for construction site management, but such systems do not appear to be widely available today or capable of coordinating human and robotic activity.

Having all these kinds of logistical things run together really well, its something I think AI can do. But its definitely going to take some time for the whole orchestration to be done well, for the right materials to get to the right place at the right time for the robot to pick it up and then to do the work or react if some of the material gets damaged, Man said. In the current construction methodology, its all about managing surprises, and there are millions of them happening over the course of the whole construction plan, so being able to effectively manage those exceptions is going to be a challenge.

Boston Dynamics, known for years as the maker of cutting-edge robots, also entered construction sites last year as part of its transition from an R&D outfit to a commercial company.

Like Sunflower Labs drones, Boston Dynamics four-legged Spot with a robotic grasping arm acts as a sensor platform for 360-video surveys of construction projects. Capable of climbing stairs, opening doors, and regaining its balance, the robot can also be equipped with other sensors to track progress and perform services that rely on computer vision.

An event held by TechCrunch at the University of California, Berkeley last month was one of the first opportunities Bay Area roboticists have had to convene since the pandemic precipitated an impending recession. Investors focused on robotics for industrial or agricultural settings urged startups to raise money now if they could, to be careful about costs, and to continue progress toward demonstrating product-market fit.

Speaking on a panel that included Built Robotics CEO Ready-Campbell, startups debated whether there will be a dominant platform for construction robotics. Contrary to others on the panel, Boston Dynamics construction technologist Brian Ringley said he believes platforms will emerge to coordinate multiple machines on construction sites.

I think long-term there will be enough people in the markets that there will be more competition, but ultimately its the same way we use lots of different people and lots of machines on sites now to do these things. I do believe there will be multiple morphologies on construction sites and it will be necessary to work together, Ringley said.

Tessa Lau is cofounder and CEO of Dusty Robotics, a company that makes an automated building layout called FieldPrinter. She said theres a huge opportunity for automation and human labor augmentation in an industry that currently has very little automation. Systems may emerge that are capable of doing the work of multiple trades or on-site activity management, but Lau said there can be nearly 80 different building trades involved in a construction site. Another problem: Construction sites are by definition in various stages of fairly constant change. The dynamic nature of construction sites where there is no set or static state like you might find in a factory presents another challenge.

I think the flip side is if you look at a typical construction site, its chaos, and anyone with a robotics background who knows anything about robotics knows its really hard to make robots work in that kind of unstructured environment, she said.

One thing the TechCrunch panelists agreed on is that robots on construction sites wont succeed unless the people working alongside them want them to. To help ensure that happens, Lau suggested startups slap googly eyes on their robots because people want to see things that are cute or beloved succeed.

Our customers are rightfully concerned that robots are going to take their jobs, and so we have to be careful about whether we are building a robot or building a tool, Lau said. And, in fact, we call our product a FieldPrinter. Its an appliance like a printer. It uses a lot of robotic technology it uses sensors and path planning and AI and all the stuff that powers robotics today, but the branding and marketing is really around the functionality. Nobody wants to buy a robot; they want to solve a problem.

Built Robotics CEO Ready-Campbell wholeheartedly agreed, arguing that even a thermostat can be considered a robot if the only requirement to meet that definition is that its a machine capable of manipulating its environment.

Last month, just before economic activity began to slow and shelter-in-place orders took effect, the International Union of Operating Engineers, which has over 400,000 members, established a multi-year training partnership with Built Robotics. Executives from Built Robotics say its systems operate primarily in rural areas that experience skilled labor shortages, but Ready-Campbell thinks its still a good idea to drop the term robot because it scares people. Opposition to construction robotics could also become an issue in areas that see high levels of unemployment.

Thats how we position Built [Robotics] in the industry, because when people think of robots, it kind of triggers a bunch of scary thoughts. Some people think about The Terminator, some people think about losing jobs, he said. Its an industry that really depends on using advanced machinery and advanced technology, and so we think that automation is just the next step in the automation of that industry.

Read more from the original source:

Robots, AI, and the road to a fully autonomous construction industry - VentureBeat

Normally, supercomputers installed at academic and national laboratories get configured once, acquired as quickly as possible before the money runs out, installed and tested, qualified for use, and put to work for a four or five or possibly longer tour of duty. It is a rare machine that is upgraded even once, much less a few times.

But that is not he case with the Corona system at Lawrence Livermore National Laboratory, which was commissioned in 2017 when North America had a total solar eclipse and hence its nickname. While this machine, procured under the Commodity Technology Systems (CTS-1) to not only do useful work, but to assess the CPU and GPU architectures provided by AMD, was not named after the coronavirus pandemic that is now spreading around the Earth, the machine is being upgraded one more time to be put into service as a weapon against the SARS-CoV-2 virus which caused the COVID-19 illness that has infected at least 2.75 million people (confirmed by test, with the number very likely being higher) and killed at least 193,000 people worldwide.

The Corona system was built by Penguin Computing, which has a long-standing relationship with Lawrence Livermore National Laboratory, Los Alamos National Laboratory, and Sandia National Laboratories the so-called Tri-Labs that are part of the US Department of Energy and that coordinate on their supercomputer procurements. The initial Corona machine installed in 2018 had 164 compute nodes, each equipped with a pair of Naples Epyc 7401 processors, which have 24 cores each running at 2 GHz with an all core turbo boost of 2.8 GHz. The Penguin Tundra Extreme servers that comprise this cluster have 256 GB of main memory and 1.6 TB of PCI-Express flash. When the machine was installed in November 2018, half of the nodes were equipped with four of AMDs Radeon Instinct MI25 GPU accelerators, which had 16 GB of HBM2 memory each and which had 768 gigaflops of FP64 performance, 12.29 teraflops of FP32 performance, and 24.6 teraflops of FP16 performance. The 7,872 CPU cores in the system delivered 126 teraflops at FP64 double precision all by themselves, and the Radeon Instinct MI25 GPU accelerators added another 251.9 teraflops at FP64 double precision. The single precision performance for the machine was obviously much higher, at 4.28 petaflops across both the CPUs and GPUs. Interestingly, this machine was equipped with 200 Gb/sec HDR InfiniBand switching from Mellanox Technologies, which was obviously one of the earliest installations of this switching speed.

In November last year, just before the coronavirus outbreak or, at least we think that was before the outbreak, that may turn out to not be the case AMD and Penguin worked out a deal to installed four of the much more powerful Radeon Instinct MI60 GPU accelerators, based on the 7 nanometer Vega GPUs, in the 82 nodes in the system that didnt already have GPU accelerators in them. The Radeon Instinct MI60 has 32 GB of HBM2 memory, and has 6.6 teraflops of FP64 performance, 13.3 teraflops of FP32 performance, and 26.5 teraflops of FP16 performance. Now the machine has 8.9 petaflops of FP32 performance and 2.54 petaflops of FP64 performance, and this is a much more balanced 64-bit to 32-bit performance, and it makes these nodes more useful for certain kinds of HPC and AI workloads. Which turns out to be very important to Lawrence Livermore in its fight against the COVID-19 disease.

To find out more about how the Corona system and others are being deployed in the fight against COVID-19, and how HPC and AI workloads are being intertwined in that fight, we talked to Jim Brase, deputy associate director for data science at Lawrence Livermore.

Timothy Prickett Morgan: It is kind of weird that this machine was called Corona. Foreshadowing is how you tell the good literature from the cheap stuff. The doubling of performance that just happened late last year for this machine could not have come at a better time.

Jim Brase: It pretty much doubles the overall floating point performance of the machine, which is great because what we are mainly running on Corona is both the molecular dynamics calculations of various viral and human protein components and then machine learning algorithms for both predictive models and design optimization.

TPM: Thats a lot more oomph. So what specifically are you doing with it in the fight against COVID-19?

Jim Brase: There are two basic things were doing as part of the COVID-19 response, and this machine is almost entirely dedicated to this although several of our other clusters at Lawrence Livermore are involved as well.

We have teams that are doing both antibody and vaccine design. They are mainly focused on therapeutic antibodies right now. They are basically designing proteins that will interact with the virus or with the way the virus interacts with human cells. That involves hypothesizing different protein structures and computing what those structures actually look like in detail, then computing using molecular dynamics the interaction between those protein structures and the viral proteins or the viral and human cell interactions.

With this machine, we do this iteratively to basically design a set of proteins. We have a bunch of metrics that we try to optimize on binding strength, the stability of the binding, stuff like that and then we do a detailed molecular dynamics calculations to figure out the effective energy of those binding events. These metrics determine the quality of the potential antibody or vaccine that we design.

TPM: To wildly oversimplify, this SARS-CoV-2 virus is a ball of fat with some spikes on it that wreaks havoc as it replicates using our cells as raw material. This is a fairly complicated molecule at some level. What are we trying to do? Stick goo to it to try to keep it from replicating or tear it apart or dissolve it?

Jim Brase: In the case of in the case of antibodies, which is what were mostly focusing on right now, we are actually designing a protein that will bind to some part of the virus, and because of that the virus then changes its shape, and the change in shape means it will not be able to function. These are little molecular machines that they depend on their shape to do things.

TPM: Theres not something that will physically go in and tear it apart like a white blood cell eats stuff.

Jim Brase: No. Thats generally done by biology, which comes in after this and cleans up. What we are trying to do is what we call neutralizing antibodies. They go in and bind and then the virus cant do its job anymore.

TPM: And just for a reference, what is the difference between a vaccine and an antibody?

Jim Brase: In some sense, they are the opposite of each other. With a vaccine, we are putting in a protein that actually looks like the virus but it doesnt make you sick. It stimulates the human immune system to create its own antibodies to combat that virus. And those antibodies produced by the body do exactly the same thing we were just talking about Producing antibodies directly is faster, but the effect doesnt last. So it is more of a medical treatment for somebody who is already sick.

TPM: I was alarmed to learn that for certain coronaviruses, immunity doesnt really last very long. With the common cold, the reason we get them is not just because they change every year, but because if you didnt have a bad version of it, you dont generate a lot of antibodies and therefore you are susceptible. If you have a very severe cold, you generate antibodies and they last for a year or two. But then youre done and your body stops looking for that fight.

Jim Brase: The immune system is very complicated and for some things it creates antibodies that remembers them for a long time. For others, its much shorter. Its sort of a combination of the of the what we call the antigen the thing about that, the virus or whatever that triggers it and then the immune system sort of memory function together, cause the immunity not to last as long. Its not well understood at this point.

TPM: What are the programs youre using to do the antibody and protein synthesis?

Jim Brase: We are using a variety of programs. We use GROMACS, we use NAMD, we use OpenMM stuff. And then we have some specialized homegrown codes that we use as well that operate on the data coming from these programs. But its mostly the general, open source molecular mechanics and molecular dynamics codes.

TPM: Lets contrast this COVID-19 effort with like something like SARS outbreak in 2003. Say you had the same problem. Could you have even done the things you are doing today with SARS-CoV-2 back then with SARS? Was it even possible to design proteins and do enough of them to actually have an impact to get the antibody therapy or develop the vaccine?

Jim Brase: A decade ago, we could do single calculations. We could do them one, two, three. But what we couldnt do was iterate it as a design optimization. Now we can run enough of these fast enough that we can make this part of an actual design process where we are computing these metrics, then adjusting the molecules. And we have machine learning approaches now that we didnt have ten years ago that allow us to hypothesize new molecules and then we run the detailed physics calculations against this, and we do that over and over and over.

TPM: So not only do you have a specialized homegrown code that takes the output of these molecular dynamics programs, but you are using machine learning as a front end as well.

Jim Brase: We use machine learning in two places. Even with these machines and we are using our whole spectrum of systems on this effort we still cant do enough molecular dynamics calculations, particularly the detailed molecular dynamics that we are talking about here. What does the new hardware allow us to do? It basically allows us to do a higher percentage of detailed molecular dynamics calculations, which give us better answers as opposed to more approximate calculations. So you can decrease the granularity size and we can compute whole molecular dynamics trajectories as opposed to approximate free energy calculations. It allows us to go deeper on the calculations, and do more of those. So ultimately, we get better answers.

But even with these new machines, we still cant do enough. If you think about the design space on, say, a protein that is a few hundred amino acids in length, and at each of those positions you can put in 20 different amino acids, you on the order of 20200 in the brute force with the possible number of proteins you could evaluate. You cant do that.

So we try to be smart about how we select where those simulations are done in that space, based on what we are seeing. And then we use the molecular dynamics to generate datasets that we then train machine learning models on so that we are basically doing very smart interpolation in those datasets. We are combining the best of both worlds and using the physics-based molecular dynamics to generate data that we use to train these machine learning algorithms, which allows us to then fill in a lot of the rest of the space because those can run very, very fast.

TPM: You couldnt do all of that stuff ten years ago? And SARS did not create the same level of outbreak that SARS-CoV-2 has done.

Jim Brase: No, these are all fairly new early new ideas.

TPM: So, in a sense, we are lucky. We have the resources at a time when we need them most. Did you have the code all ready to go for this? Were you already working on this kind of stuff and then COVID-19 happened or did you guys just whip up these programs?

Jim Brase: No, no, no, no. Weve been working on this kind of stuff for her for a few years.

TPM: Well, thank you. Id like to personally thank you.

Jim Brase: It has been an interesting development. Its both been both in the biology space and the physics space, and those two groups have set up a feedback loop back and forth. I have been running a consortium called Advanced Therapeutic Opportunities in Medicine, or ATOM for short, to do just this kind of stuff for the last four years. It started up as part of the Cancer Moonshot in 2016 and focused on accelerating cancer therapeutics using the same kinds of ideas, where we are using machine learning models to predict the properties, using both mechanistic simulations like molecular dynamics, but all that combined with data, but then also using it other the other way around. We also use machine learning to actually hypothesize new molecules given a set of molecules that we have right now and that we have computed properties on them that arent quite what we want, how do we just tweak those molecules a little bit to adjust their properties in the directions that we want?

The problem with this approach is scale. Molecules are atoms that are bonded with each other. You could just take out an atom, add another atom, change a bond type, or something. The problem with that is that every time you do that randomly, you almost always get an illegal molecule. So we train these machine learning algorithms these are generative models to actually be able to generate legal molecules that are close to a set of molecules that we have but a little bit different and with properties that are probably a little bit closer to what we what we want. And so that allows us to smoothly adjust the molecular designs to move towards the optimization targets that we want. If you think about optimization, what you want are things with smooth derivatives. And if you do this in sort of the discrete atom bond space, you dont have smooth derivatives. But if you do it in these, these are what we call learned latent spaces that we get from generative models, then you can actually have a smooth response in terms of the molecular properties. And thats what we want for optimization.

The other part of the machine learning story here is these new types of generative models. So variational autoencoders, generative adversarial models the things you hear about that generate fake data and so on. Were actually using those very productively to imagine new types of molecules with the kinds of properties that we want for this. And so thats something we were absolutely doing before COVID-19 hit. We have taken these projects like ATOM cancer project and other work weve been doing with DARPA and other places focused on different diseases and refocused those on COVID-19.

One other thing I wanted to mention is that we havent just been applying biology. A lot of these ideas are coming out of physics applications. One of our big things at Lawrence Livermore is laser fusion. We have 192 huge lasers at the National Ignition Facility to try to create fusion in a small hydrogen deuterium target. There are a lot of design parameters that go into that. The targets are really complex. We are using the same approach. Were running mechanistic simulations of the performance of those targets, we are then improving those with real data using machine learning. So now we now have a hybrid model that has physics in it and machine learning data models, and using that to optimize the designs of the laser fusion target. So thats led us to a whole new set of approaches to fusion energy.

Those same methods actually are the things were also applying to molecular design for medicines. And the two actually go back and forth and sort of feed on each other and support each other. In the last few weeks, some of the teams that have been working on the physics applications have actually jumped over onto the biology side and are using some of the same sort of complex workflows that were using on these big parallel machines that theyve developed for physics and applying those to some of the biology applications and helping to speed up the applications on these on this new hardware thats coming in. So it is a really nice synergy going back and forth.

TPM: I realize that machine learning software uses the GPUs for training and inference, but is the molecular dynamics software using the GPUs, too?

Jim Brase: All of the molecular dynamics software has been set up to use GPUs. The code actually maps pretty naturally onto the GPU.

TPM: Are you using the CUDA variants of the molecular dynamics software, and I presume that it is using the Radeon Open Compute, or ROCm, stack from AMD to translate that code so it can run on the Radeon Instinct accelerators?

Jim Brase: There has been some work to do, but it works. Its getting its getting to be pretty solid now, thats one of the reasons we wanted to jump into the AMD technology pretty early, because you know, any time you do first-in-kind machines its not always completely smooth sailing all the way.

TPM: Its not like Lawrence Livermore has a history of using novel designs for supercomputers. [Laughter]

Jim Brase: We seldom work with machines that are not Serial 00001 or Serial 00002.

TPM: Whats the machine learning stack you use? I presume it is TensorFlow.

Jim Brase: We use TensorFlow extensively. We use PyTorch extensively. We work with the DeepChem group at Stanford University that does an open chemistry package built on TensorFlow as well.

TPM: If you could fire up an exascale machine today, how much would it help in the fight against COVID-19?

Jim Brase: It would help a lot. Theres so much to do.

I think we need we need to show the benefits of computing for drug design and we are concretely doing that now. Four years ago, when we started up ATOM, everybody thought this was nuts, the general idea that we could lead with computing rather than experiment and do the experiments to focus on validating the computational models rather than the other way around. Everybody thought we were nuts. As you know, with the growth of data, the growth of machine learning capabilities, more accessibility to sophisticated molecular dynamics, and so on its much more accepted that computing is a big part of this. But we still have a long way to go on this.

The fact is, machine learning is not magic. Its a fancy interpolator. You dont get anything new out of it. With the physics codes, you actually get something new out of it. So the physics codes are really the foundation of this. You supplement them with experimental data because theyre not right necessarily, either. And then you use the machine learning on top of all that to fill in the gaps because you havent been able to sample that huge chemical and protein space adequately to really understand everything at either the data level or the mechanistic level.

So thats how I think of it. Data is truth sort of and what you also learn about data is that it is not always the same as you go through this. But data is the foundation. Mechanistic modeling allows us to fill in where we just cant measure enough data it is too expensive, it takes too long, and so on. We fill in with mechanistic modeling and then above that we fill in that then with machine learning. We have this stack of experimental truth, you know, mechanistic simulation that incorporates all the physics and chemistry we can, and then we use machine learning to interpolate in those spaces to support the design operation.

For COVID-19, there are there are a lot of groups doing vaccine designs. Some of them are using traditional experimental approaches and they are making progress. Some of them are doing computational designs, and that includes the national labs. Weve got 35 designs done and we are experimentally validating those now and seeing where we are with them. It will generally take two to three iterations of design, then experiment, and then adjust the designs back and forth. And were in the first round of that right now.

One thing were all doing, at least on the public side of this, is we are putting all this data out there openly. So the molecular designs that weve proposed are openly released. Then the validation data that we are getting on those will be openly released. This is so our group working with other lab groups, working with university groups, and some of the companies doing this COVID-19 research can contribute. We are hoping that by being able to look at all the data that all these groups are doing, we can learn faster on how to sort of narrow in on the on the vaccine designs and the antibody designs that will ultimately work.

Excerpt from:

One Supercomputers HPC And AI Battle Against The Coronavirus - The Next Platform

Now is the age of home entertainment. The concept of the modern home goes way beyond being merely a residential space; it has become a place for relaxation, for recreation and for quality time with others. At the center of this change is the near-three-dimensional content experiences granted by todays ultra-large, ultra-high definition, ultra-fine pixel TVs. Of course, high quality audio provides the finishing touch to such experiences.

With its 2020 QLED 8K TVs and AI sound technologies, Samsung has raised the bar for TV audio experiences. Object Tracking Sound+ uses AI-based software to match the movement of audio with movement on-screen; Active Voice Amplifier (AVA) tracks the users audio environment; and Q-Symphony creates a more realistic, three-dimensional sound.

Samsung Newsroom sat down with the sound developers of Samsung Electronics Visual Display Business to learn more about their extensive capabilities and the journey to fostering innovation in sound.

(From left to right) Youngtae Kim (Sound Lab), Jongbae Kim (Sound Lab), Yoonjae Lee (Sound Device Lab) and Sunmin Kim (Sound Lab)

Action movies with amazing sound arrangements provide the most realistic experiences when experienced at movie theaters. This is because movie theaters have multi-channel speakers with 3D sound placed on almost all the walls (including the ceiling), as well as around the screen. Compared with two-channel sound that features speakers only on the left- and right-hand sides, the multi-channel speakers in a movie theater deliver a more refined sense of realism. So how can this realism be recreated in the home? Samsungs sound developers came up with Object Tracking Sound+ technology, in which sound follows movement onscreen through six speakers built into the TV.

Thanks to this technology, a videos audio follows the action on-screen in real time. When a car moves from the left to the right-hand side of the screen, so will the sound it makes; and when a heavy object drops from the top to the bottom of the screen, so will the audio.

When developing the 2020 QLED 8K TVs, Samsungs TV developers increased the number of QLED TV speakers from two to six in order to realize sound that can mimic action. By placing two speakers on each side of the screen, as well as on the top and bottom, we enabled the dissemination of sound in all directions from a total of six speakers, explained Jongbae Kim. The distance between the two main speakers has been widened as much as possible, and the additional speakers have been installed in order to maximize sound across all axes to be as three-dimensional as possible. For example, we placed speakers on the upper side of the screen to enable the movement of sound in a vertical direction for a more immersive sound experience. Additionally, Kim highlighted how, despite the complex nature of a TV structure that includes six embedded speakers, the team managed to keep the design of the TV slim and minimal.

In order to ensure audio will follow on-screen movements accurately, it is important to understand the original intentions of content creators. The role of a sound engineer is to increase the consistency between the action onscreen and its accompanying audio track when mixing. The location information of sound in a piece of video content, including sound panning information, is subsequently audio-signaled into the audio channel by the sound engineer something we must be able to track in order to reproduce the location and movement of a contents audio accurately, noted Jongbae Kim. Our Object Tracking Sound+ technology analyzes the location information contained in these audio signals as originally placed during mixing. This means the TV can then effectively distribute the sound amongst its six speakers by distinguishing between sound orientations and whether or not the audio source is on-screen, off-screen, close-up or distant.

When that crucial scene in the show you are watching is overwhelmed by mixer sound or an important breaking news report is obscured by loud thunder, the act of reaching for the remote to adjust the TVs volume can come too late. This is why the team developed their AVA technology, which recognizes exterior noises and increases the volume of voices in content accordingly if surrounding conditions become too loud.

The way it works is intelligent. The TVs sound sensor, attached to the bottom middle of the TV, takes in audio from the content onscreen as well as its surrounding environs. AI technology then compares the volume levels of the two types of sound, and if external sounds are found to be louder than that of the TVs content, it will selectively raise the TVs volume. The system does not have one set definition of noise, explained Sunmin Kim. It considers any and all elements that disturb enjoyment of content as noise. When exterior sounds persist above a certain decibel level, that is when the system registers it as noise.

However, AVA technology does not just raise the volume of the TV when it recognizes a louder environment, as this would only contribute to a boisterous room experience. The system harnesses AI to keep sound effect and background audio levels consistent and to only raise the volume of voice audio, highlighted Sunmin Kim. Our research showed us that the majority of content is dialogue-heavy, so we believe that enhancing the delivery of dialogue would be most beneficial to aid comprehension.

One of key elements to achieving a realistic sound during content playback on a TV is three-dimensionality, which encompasses both horizontal and vertical audio characteristics. Until recently, these perspectives had been being developed separately by the TV and Soundbar teams. However, with the inclusion of upper side speakers on Samsungs 2020 QLED 8K TVs, the team developed an all-inclusive solution that utilizes the capabilities of both the TV and the soundbar in perfect harmony. Q-Symphony is a feature that plays audio using both the TV speakers and the soundbar at the same time, and as an industry first achievement, the Q-Symphony technology was even recognized with a Best of Innovation Award at CES 2020.

The core of Q-symphony, which manages sound playback harmoniously using speakers with different characteristics, is technology that follows the sound playback rules which are determined in advance and exchanges the necessary information between the TV and the soundbar, when connected. This approach allows for a superior reproduced sound experience, explained Yoonjae Lee. A key element of the technology is a special algorithm that we created which divides and harmonizes sounds seamlessly between the TVs speakers and soundbar.

During development, challenge arose regarding the quality of dialogue reproduction. When both the soundbar and the TV speakers played dialogue simultaneously, the sound quality was noticeably diminished. However, the sound development team were able to resolve this issue by separating the main sound track, including dialogue, from the entire signal and assigning the different tracks to the TV speakers and soundbar respectively. In the 2020 QLED 8K TV range, the voice signal is extracted and removed from the sound being reproduced by the TVs embedded speakers, which are then assigned to play ambient sound signals such as sound effects, explained Lee. The soundbar then reproduces the main sound involving any dialogue. With this technology, Q-Symphony harnesses the advantages of both TV speakers and the soundbar in order to deliver the best, and most harmonious, sound experience to users.

The sound development team agreed that realizing the addition of speakers, the new placement of speakers and the AI harmonization with the soundbar on the 2020 QLED 8K TV range was possible because of close coordination with a variety of other teams. When developing new TVs, all areas need to be in sync with their innovations, noted Youngtae Kim. We came together in suggesting various solutions to overcome each and every technological hurdle with an open mind.

Youngtae Kim (left) and Sunmin Kim introduce the innovative sound technologies of the 2020 QLED 8K TV range

The sound development team has always been and always will be dedicated to developing the best possible audio experiences for users. As well as working with Samsungs Audio Lab based in the U.S. for the future audio technology, the team also works with Samsung Researchs R&D Centers, universities and start-up experts around the world. We want to bring about sound experiences that are as natural and as real as possible, explained Youngtae Kim. To achieve this, we will continue to work hard to understand the end-to-end process of the sound and realize sounds in our TVs.

Samsungs sound innovation also helps to realize its vision of Screens Everywhere. In the future, we will bolster the use of AI so that users do not even need to use a remote control to find the perfect audio balance when enjoying their favorite content, said Sunmin Kim. As time goes by, TV installation environments, lifestyles and age groups will be diversified. We want users to enjoy the sound of their content as intended, regardless of content type or listening environment.

Read more from the original source:

[The Future of Viewing] Innovative Sound Technologies, Powered by AI - Samsung Global Newsroom

Article By : Geoff Tate

Inference chips typically have lots of MACs and memory but actual throughput on real-world models is often lower than expected. Software is usually the culprit.

Inference accelerators represent an incredible market opportunity not only to chip and IP companies, but also to the customers who desperately need them. As inference accelerators come to market, a common comment we hear is: Why is my inference chip not performing like it was designed to?

Oftentimes, the simple answer is the software.

Software is keyAll inference accelerators today are programmable because customers believe their model will evolve over time. This programmability will allow them to take advantage of enhancements in the future, something that would not be possible with hard-wired accelerators. However, customers want this programmability in a way where they can get the most throughput for a certain cost, and for a certain amount of power. This means they have to use the hardware very efficiently. The only way to do this is to design the software in parallel with the hardware to make sure they work together very well to achieve the maximum throughput.

One of the biggest problems today is that companies find themselves with an inference chip that has lots of MACs and tons of memory, but actual throughput on real-world models is lower than expected because much of the hardware is idling. In almost every case, the problem is that the software work was done after the hardware was built. During the development phase, designers have to make many architectural tradeoffs and they cant possibly do those tradeoffs without working with both the hardware and software and this needs to be done early on. Chip designers need to closely study the models, and then build a performance estimation model to determine how different amounts of memory, MACs, and DRAM would change relevant throughput and die size; and how the compute units need to coordinate for different kinds of models.

Today, one of the highest-volume applications for inference acceleration is object detection and recognition. That is why inference accelerators must be very good at mega-pixel processing using complex algorithms like YOLOv3. To do this, it is critical that software teams work with hardware teams throughout the entire chip design process from performance estimation to building the full compiler and when generating code. As the chip designer has the chip RTL done, the only way to verify the chip RTL at the top level is to run entire layers of models through the chip with mega-pixel images. You need to have the ability to generate all the code (or bit streams) that control the device and that can only be done when software and hardware teams work closely together.

Today, customer models are neural networks and they come in ONNX or TensorFlow Lite. Software takes these neural networks and applies algorithms to configure the interconnect and state machines that control the movement of data within the chip. This is done in RTL. The front end of the hardware is also written in RTL. Thus, the engineering team that is writing the front-end design is talking a similar language to the people that are writing the software.

Why software also matters in the futureFocusing on software is not only critical early on, but will also be critical in the future. Companies that want to continue delivering improvements are going to need their hardware teams studying how the software is evolving and how the models emerging are shifting in a certain direction. This will enable chip designers to make changes as needed, while the company also improves their complier and algorithms to better utilize the hardware over time.

In the future, we expect companies to continue bringing very challenging models to chip designers, with the expectation that new inference accelerators can deliver the performance needed to handle those models. Like we see today, many chip companies may try and cut costs and development times by focusing more on the hardware initially. However, when the chips are done and delivered to market, its going to be the ones that focused on software early on that will offer the best performance and succeed.

Geoff Tate is CEO of Flex Logix Technologies

Related Posts:

See original here:

Why Software Matters in the Age of AI? - Eetasia.com

April10, 20205 min read

Opinions expressed by Entrepreneur contributors are their own.

In only a few weeks, the COVID-19 pandemic has completely disrupted our normal way of life. With many businesses shutting their doors or transitioning to a work-from-home system, adaptability to a constantly changing situation will prove key for the survival of organizations large and small. Despite everything that is going on, however, the pandemic is also spurring new innovations, particularly in the world of artificial intelligence. Here are several important ways AI is already making a difference in improvingpublic health and safety as the world adapts to a new normal.

One of the biggest challenges with this coronavirus (and the COVID-19 disease it subsequenly causes) has been how quickly it can spread. While social-distancing measures and the closure of high-risk facilities are viewed as the best way to control the spread, many areas have been slow to enact such measures because they dont have an accurate perception of their risk.

In Israel, however, an AI-powered survey system developed by the Weizmann Institute of Science aims to better predict outbreaks so authorities can proactively enact measures that will mitigate the viruss spread. The system uses a questionnaire focusing on key issues like health symptoms and isolation practices, then matches responses with a location-based algorithm. AI analysis can then identify potential hotspots in advance, which can help local authorities enact measures that will slow down the virus.

With COVID-19 constantly dominating headlines, it should come as no surprise that hospitals and health organizations are getting more inquiries than ever from patients worried that they might have the coronavirus.

Virtual assistants have already alleviated the workloads of customer support professionals in other industries, and now, similar tools specifically designed to address questions related to COVID-19 are being introduced. These AI tools can be embedded directly into healthcare apps and websites.

One example of this is Hyro, a free COVID-19 virtual assistant that is being offered to healthcare organizations to help them manage the uptick in calls and questions. By answering frequently asked questions about the coronavirus, triaging symptoms and delivering information from verified sources like the WHO and CDC, such AI tools can help reduce the burden on healthcare workers who are already being stretched thin by pandemic conditions.

Related: How Artificial Intelligence Is Helping Fight the COVID-19 Pandemic

An unfortunate issue that has popped up in the wake of the COVID-19 pandemic is the rapid spread of misinformation online. From downplaying the risks posed by the virus to false text messages warning of mandatory quarantine orders, this can further fuel panic during what is already a scary time.

Many social media platforms use human content moderators to check for harmful posts, but with more employees being required to work from home or stop working altogether, AI is becoming increasingly important in combating misinformation. Though the lack of human supervision means an increased risk for mistakes, it could also spur new improvements for these machine-learning tools.

As one example of this, The Verges Jacob Kastrenakes explains, YouTube will rely more on AI to moderate videos during the coronavirus pandemic, since many of its human reviewers are being sent home to limit the spread of the virus. This means videos may be taken down from the site purely because theyre flagged by AI as potentially violating a policy, whereas the videos might normally get routed to a human reviewer to confirm that they should be taken down.

As noted by the Guardian, one of the biggest challenges in containing the spread of COVID-19 is the fact that many patients experience symptoms most similar to a mild cold. Some are entirely asymptomatic. Because of this, many people who could spread the virus to others may continue to go out in public rather than self-quarantining.

While testing can be slow, AI is already stepping up to the challenge. As reported by The Next Web, several AI tools have already been developed to identify patients with COVID-19 and deliver treatment that keeps healthcare professionals safe.

In China, a computer-vision algorithm was developed to scan peoples temperatures in public locations and flag anyone with even a slight fever. Another AI algorithm helps doctors more accurately discern between coronavirus and typical pneumonia patients. In Washington State, robots have even been used to provide remote treatment and communication to keep the disease from spreading from patients to doctors.

Related: How Businesses Should Handle the Coronavirus Outbreak

The future surrounding the COVID-19 pandemic is rife with uncertainty. There is no telling how long social isolation measures and other precautions will need to remain in place to mitigate the spread of the disease, or what the overall impact of such actions will be.

While AI may not have all the answers, it is clear that continuing innovation in this field will help and already is helping to make the world a safer place during these troubling times. By helping slow the spread of the virus and improving conditions for healthcare workers, these tech developments could very well save lives now and in the future.

Read the original:

4 Ways AI Is Making the World a Safer Place - Entrepreneur

TensTorrent founder and CEO Ljubisa Bajic talking remotely at the Linley Group Spring Processor Forum.

2016 was an amazing year in the history of computing. That year, numerous experienced computer chip designers set out on their own to design novel kinds of parts to improve the performance of artificial intelligence.

It's taken a few years, but the world is finally seeing what those young hopefuls have been working on. The new chips coming out suggest, as ZDNet has reported in past, that AI is totally changing the nature of computing. It also suggests that changes in computing are going to have an effect on how artificial intelligence programs, such as deep learning neural networks, are designed.

Case in point, startup Tenstorrent, founded in 2016 and headquartered in Toronto, Canada, on Thursday unveiled its first chip, "Grayskull," at a microprocessor conference run by the legendary computer chip analysis firm The Linley Group. The "Spring Processor Conference" was originally going to be held in Silicon Valley, but COVID-19 turned it into a Zoom video affair. The event was very well attended, organizer Linley Gwennap told ZDNet, and interest was evident from the number of audience questions submitted via Zoom chats.

Tenstorrent was founded by chief executive Ljubisa Bajic, who previously worked for Nvidia and Advanced Micro Devices, among others. As Bajic related to ZDNet, in 2016 he was trying to figure out his next move after several years at those large chip vendors. A legendary chip designer, Jim Keller, who worked with Bajic at AMD, "told me to just go and do what interested me," said Bajic. Keller wrote a check to Bajic, providing the first funding for the company.

The company has now received a total of $34 million in funding from Eclipse Ventures and Real Ventures, among others. The company also has offices in Austin, Texas and in Silicon Valley.

Bajic, and other chip teams, are responding to the explosion in the size of deep learning models, such as BERT, and OpenAI's "GPT2," but also even newer models such as Nvidia's "Megatron," Microsoft's "Turing NLG," and neural net models that Bajic said he couldn't talk about publicly that will have on the order of one-trillion parameters.

The Grayskull chip is meant to be used to speed up what's called "inference," the part of AI where a trained neural network makes predictions from new data. This part of the market has traditionally been dominated by Intel microprocessors in server computers in data centers. But Nvidia has made big inroads into inference with its graphics processing units (GPUs), and numerous startup companies have announced chip designs to compete with both of those chip giants.

The chip is expected to go into production this fall, Bajic told the conference.

The only way to beat Nvidia is with vastly superior performance, Bajic told ZDNet in an interview. "Most customers are not going to switch off of Nvidia for a part that is only two times better that is basically still an engineering sample," said Bajic. "Our goal is if we can be more than ten times better than Nvidia and sustain that for a few years, then we think people will come around to us if we can achieve that."

Early results look good. In a review of the Grayskull part put out Thursday, the Linley Group's lead analyst. Linley Gwennap, writes that the Grayskull part has "excellent" performance relative to Nvidia and other startups, including a band of former Google engineers named Groq. In fact, the chip is more efficient at performing standard AI tasks than all other chips on the market, including Nvidia's, leaving aside processors that Chinese search giant Alibaba uses internally. For example, the chip can perform 368 trillion "operations per second" on a circuit board consuming just 75 watts versus parts from Nvidia and others that require 300 watts on average. (Subscription required to read Linley Group articles.)

The "Grayskull" 75-watt PCIe card.

What's going on in the Grayskull chip has interesting implications for computing. One focus is lots and lots of computer memory. The Grayskull part has 120 megabytes of on-chip SRAM memory, compared to just 18 megabytes for Nvidia's "Titan RTX" part. Nvidia's approach has been to hook up its GPUs to the fastest off-chip memory. But Tenstorrent and other startups are increasingly emphasizing the role of faster on-chip memory.

For example, Groq's "TSP" chip has almost twice as much memory as Tenstorrent, at 220 megabytes. And the record for on-chip memory is held by Cerebras Systems, which also presented at Thursday's conference. Cerebras's part, the world's largest chip, called the "Wafer Scale Engine," which was unveiled in August, has a grand total of 18 gigabytes of on-chip memory.

The proliferation of memory as a larger and larger influence in the design of the chip has some startling implications for computer system design. For example, Groq's TSP has no DRAM interface. Instead, there are several connectors around the edges of the chip that are called "SERDES" that are the kind of connectors that are used in data networking. The idea, explained Dennis Abts, who spoke following Bajic, is that instead of adding external DRAM, one can combine multiple Groq chips together through the SERDES, so that all memory operations are handled by multiplying the available on-chip SRAM, with no DRAM whatsoever. Like Cerebras and Tenstorrent, and other companies such as Graphcore, the ultimate vision for Groq is that people will use many of its chips in massively parallel computers that are combine multiple boards together. Hence, the era of external DRAM may be drawing to a close, replaced by on-chip SRAM in massively parallel computers.

As far as speeding up AI, having lots of on-chip memory fulfills a bunch of functions. One is to keep memory close to the multiple on-chip computing cores. Tenstorrent has 120 computing cores on Grayskull, and Cerebras has 400,000 compute cores. The large on-chip memory is spread amongst these cores; it resides in the circuitry that is closest to the computing core, so that it takes no more than a single tick of the chip's clock for each core to access the memory it needs to read or write.

A schematic of the Grayskull chip, with its 120 on-chip compute cores.

As Cerebras's head of hardware engineering, Sean Lie, noted, "In machine learning, the weights and the activations are local, and there's low data reuse," he noted. "But the traditional memory hierarchy isn't built that way." Instead, general-purpose chips like Intel Xeon CPUs and GPUs spend a lot of time going all the way off the chip to external DRAM memory, which takes several clock cycles to access. By keeping the values that any one processing core is working on, instead of going away to off-chip DRAM, "the physics of local memory drives higher performance," said Lie.

There is a secondary efficiency, notes Lie, which is reducing the duplication that happens from using off-chip memory. GPUs "turn vector-matrix multiplies into matrix-matrix multiplies," he said, meaning, they bunch several inputs together, what's known as "batching," which is the bte noir of most deep learning scientists. "It actually changes the training of machine learning," distorting results, said Lie. "Remove the large-batch multiplier is a goal," said Lie. That point was echoed by Tenstorrent's Bajic. "No more general matrix multiplication," Bajic told the conference, "No more batching."

The kinds of neural net models that AI chips have to plan to handle, especially in the domain of natural language, are scaling to very, very large numbers of parameters, over a trillion, argues Tenstorrent founder and CEO Ljubisa Bajic.

Instead of batching, all three companies, Tenstorrent, Groq, and Cerebras, emphasize "sparsity," where individual inputs to a neural network are treated independently by individual cores on the chip. That's where the implications become very interesting for machine learning.

As they dispense with batching, Tenstorrent and the other companies are fixated on the concept of "sparsity," the notion that many neural networks can be processed more efficiently if redundant information is stripped away. Lie observed that there is "a large, untapped potential for sparsity" and that "neural networks are naturally sparse."

Tenstorrent's Bajic told the conference that sparsity is at the heart of the Grayskull chip. One of the big influences upon the chip's design is the way that the brain's neurons only fire some of the time they spike. Much of the time, those neurons are idle, consuming little power.

"Spiking neural nets are more efficient, in a sense," Bajic said. "They have conditional execution, they have branching, etc. They are not efficient in trillions of operations per watt and terabytes per second, but they have this very nice feature of implementing functionality only when needed, and as a result they have very good power efficiency."

Tenstorrent says that with a little bit of extra training, a neural network such as Google's BERT can be optimized to exit its programming tasks "early," saving on compute effort.

Hence, the Tenstorrent team has come up with a way to streamline the way a neural network is computed in silicon to be more selective, what he called "conditional execution." A neural network, including a popular natural language program, such as Google's BERT, goes through several layers of processing. It's possible to stop that neural network before it goes through all the layers, and avoid some computation, said Bajic. By "testing" if a neural network has reached a satisfactory answer mid-way through its computation, the program can be stopped early, what's known as "early exit."

That's what Bajic and team have done, designed a software program that re-trains a neural network to find the places where it might be able to stop early. "Take BERT as a pre-packaged bunch of code, Python code calling PyTorch primitives, and add a bit of code that attaches early exit," explained Bajic. "And then run a little bit of fine-tuning training, about a half hour of additional training."

The training step goes through the entire process of compute, said Bajic, but the trained model, once ready to perform inference, can stop where it reaches a sufficient prediction, and save some compute. "There will be a statement that says, if my input is high confidence, and based on existing data, it's just not going to run rest of the network." Tenstorrent customers can either run the extra half hour of training before compiling their neural networks, or the Tenstorrent software can automatically re-train the model. "You can take networks like BERT and GPT2 and run them through our black box and get all this done so you as a user do not have to assign engineers to do it, you don't have to negotiate with the machine learning team to get it done," said Bajic.

The result of tricks such as early exit can be a dramatic speed up in performance. The Grayskull chip can process 10,150 sentences per second with BERT versus the customary 2,830 sentences that most chips can process in that time.

That's a neat trick, but it's also a change to the way that neural nets are thought about. The Grayskull part signals that what a chip can do may change how such networks are designed in future. It's like what Facebook's head of AI, Yann LeCun, pointed out a year ago: "Hardware capabilities and software tools both motivate and limit the type of ideas that AI researchers will imagine and will allow themselves to pursue."

"The tools at our disposal fashion our thoughts more than we care to admit," LeCun has said.

What is coming into focus, then, is a world in which both computing approaches and artificial intelligence approaches will be changing simultaneously, affecting one another in a symbiotic way. That means it will become increasingly difficult to talk about things such as how fast a given machine learning model runs, or how fast a chip is, without considering how different both are from past efforts. Metrics in either case will be intimately bound up with the choices made by chip designers, computer designers, and AI scientists who build neural networks.

"MLPerf has not comprehended this kind of approach to the problem at all," said Bajic, referring to benchmark chip tests. "We sort of make it not quite BERT," he said. "that's something I would invite MLPerf people to think about."

See more here:

Startup Tenstorrent shows AI is changing computing and vice versa - ZDNet

McGill University researchers say theyve developed a technique to train a remote-controlled, offroad car to drive on terrain from aerial and first-person imagery. Their hybrid approach accounts for terrain roughness and obstacles using on-board sensors, enabling it to generalize to environments with vegetation, rocks, and sandy trails.

The work is preliminary, but it might hold promise for autonomous vehicle companies that rely chiefly on camera footage to train their navigational AI. U.K.-based Wayve is in that camp, as are Tesla, Mobileye, and Comma.ai.

The researchers work combines elements of model-free and model-based AI training methods into a single graph to leverage the strength of both while offsetting their weaknesses. (As opposed to model-free methods, model-based methods have a software agent try to understand the world and create a model representing it, which sometimes leads to poor performance due to cascading errors.) Their model learns to navigate collision-free trajectories while favoring smooth terrain in a self-supervised fashion, such that the training data is labeled autonomously.

The researchers off-road vehicle is based on an electric, two-motor remote-controlled buggy with a mechanical brake thats wirelessly connected to an Intel i7 NUC computer running the open source Robot Operating System (ROS). The buggy is equipped with both a short-range lidar sensor and a forward-facing camera coupled with an inertial measurement unit, and with a microcontroller that relays all sensor information to the NUC computer.

Before deploying the buggy on an all-terrain course, the team captured images of the course from an 80-meter height using a DJI Mavic Pro, and then they extracted 12-meter-by-9-meter patches of the images so that they could be oriented and centered. The images were taken at a resolution of 0.01 meters per pixel and were aligned within 0.1 meter, using four visual landmarks measured with the buggy.

During training, the teams model estimates terrain roughness using an inertial measurement unit while the lidar sensor measures the distance between obstacles. Given fused input images from an onboard camera and local aerial view, a recent visual history, terrain class labels (e.g., rough, smooth, obstacle), and a sequence of steering commands, it predicts collision probabilities over a fixed horizon from which a policy or strategy can be derived.

In a real-world field trial, the researchers had the buggy drive at a speed of 6 kilometers per hour (~3.7 miles per hour) after training on 15,000 data samples collected over 5.25 kilometers (~3.2 miles). They report that the navigational model achieved a prediction accuracy of 60% to 78% using the forward ground camera and that when the aerial imagery was incorporated, accuracy increased by around 10% for trajectories with angle changes of 45 degrees or higher. Indeed, the policy drove on smooth terrain 90% of the time and reduced the proportion of rough terrain by over 6.1 times compared with a model using only first-person imagery.

More here:

AI training helps remote-controlled buggy negotiate rugged terrain - VentureBeat

If, like many people these days, youre trying to get a firmer understanding of what AI is and how it works but are secretly panicking a little because youre struggling with terminology so opaque that youre lost before you get to Markov chains, you may want to crack open Janelle Shanes new book. She recently sat down with VentureBeat to talk about the book, whose title, You Look Like a Thing and I Love You, is actually an AI-generated pickup line.

Shane maintains the AI Weirdness blog and combines knowledge from her Ph.D. in electrical engineering, fascination with AI, and propensity for slightly deadpan absurdist humor to explain AI in a way that is both hilarious and easy to understand. More importantly, she uses humor as a frame to display how AI is actually dangerously bad at most things we wish it could do. Her take is a refreshing counter to the often overly fantastical notions about AI, its looming sentience, and its capacity for either utopia or dystopia.

Although the book walks the reader through what AI is and explains how AI thinks and how we should think about how AI thinks, its full of giggle-inducing hand-drawn illustrations and endless comical examples. Shanes signature move is using neural nets to generate things like new ice cream flavors, knitting patterns, and recipes. The results are usually funny and bizarre and are things that could almost exist, like Lemon-Oreo ice cream or a Cluckerfluffer sandwich (chicken, cheese, and marshmallow). Its the least creepy way to show how AI so often falls off a cliff into the uncanny valley.

If you can make it past the recipe for Basic Clam Frosting without laughing (page 18), there might be something wrong with you.

Above: These flavors are not delicious. Janelle Shane at her TED Talk

Image Credit: TED

This interview has been edited for length and clarity.

VentureBeat: When I first came across [the book], it seemed like it was going to be very educational. And also very accessible. And I thought that sounded fantastic. I know this all started with AI Weirdness. Why did you start that blog? How did that come about?

Janelle Shane: I saw someone else named Tom Brewe who had used one of these text-generating neural nets to do cookbook recipes. It was one of these situations where Im laughing so hard the tears are streaming down my face. I have now read all of the recipes hes generated there now needs to be more recipes! And [I thought] I need to learn how to generate them. So thats kind of it was more like, I thought this was hilarious. I generated this thing that Im like, Okay, now I have to show people. I figured it would be a few of my friends who would read it. I had the blog from my Ph.D. days that mostly had electron microscope pictures on it. So I just threw some of these experiments on there, [thinking], you know, Heres a blog, Ill just put this stuff on there. And then, yeah, to my surprise, other people started reading it, like, people I didnt even know personally.

VentureBeat: At what point did the book emerge as a thing, coming out of the blog?

Shane: In part it was because I got [to] talking to an agent and to a publisher who were interested in there being this book. But, you know, the other motivation too was, again, repeatedly getting the comments of people [who were] confused because the AI on my blog was making all sorts of mistakes and Isnt AI supposed to be smart? So this book is kind of a way to explain how these things can be true at once and what smart means in the context of AI. And what it definitely doesnt mean.

VentureBeat: So they didnt get the joke? (Which is also kind of funny.)

Shane: It was more like they were confused, like they got that, Okay, the AI is making mistakes, and haha thats funny but why is this one so bad? And, you know, is this a different kind of AI? This is stuff that I was recommending movies or labeling photos and things.

VentureBeat: And how did you come to the illustrations? Was that kind of your thing, or did the publisher kind of come in and do that?

Shane: Those are all my own illustrations.

VentureBeat: How did you come up with that sort of language its a really funny and educational sort of artistic language. How did that come to you?

Shane: This is how I explain things to myself in my own voice. I remember kind of overhearing when I was doing rehearsals, actually, for the TED conference, and I overheard a couple of the coaches saying to each other, No, she really does sound like that! So yeah, its not an act! This is really how I talk about these things and how I like to write them down for myself so they sort of make sense. Im telling myself this story to sort of put my thoughts in order.

Definitely, this approach where focusing in on examples and especially on these kind of memorable things thats whats going to stick around after youre done reading the book. Its not like my bullet-pointed principles of AI weirdness was going to stick in [your] mind as much as the story about, you know, the AI that got confused about whether were supposed to be identifying fish and whether human fingers holding the trophy fish are part of the fish. Or, you know, the AI that was told to sort a list of numbers and deleted the list, therefore eliminating all the sorting errors, thus technically getting a perfect score.

So that sort of focusing in on these really weird, very non-human sorts of things to do I think in part comes from the first way that I encountered machine learning, which was as a freshman in college in 2002, sitting in on a lecture by a guy whos studying evolutionary algorithms. His name is Professor Erik Goodman, and he was explaining his research and telling just these same kinds of stories. And I remember that just really grabbed me and was so weird and memorable and made sense, but was also difficult to picture, you know kind of like studying the natural world in a way. And so that really kind of was the way I got drawn into machine learning.

VentureBeat: It struck me that the way you explained how neural nets work is very Amelia Bedelia-like. [Amelia Bedelia is a childrens book character who takes everything people say literally, thus exposing the humor in things like idioms.] Its almost like the math version of how Amelia Bedelia plays with language.

Shane: Yeah, you know, thats one of the things thats fun about neural nets, is that they dont have the context for the exact problem theyre trying to solve. They take things entirely literally because thats all we know how to do. And then you get some really weird situations, this really weird humor that kind of arises from that. So yeah, it definitely pushes some of the same buttons.

VentureBeat: You use real-world examples, and sort of absurdist humor, to show what happens when AI goes awry. And its kind of easy to vet the output, because it kind of comes out as gibberish, right? We can tell that the recipe doesnt make sense, and its fun to make it and sort of point that out, but I wonder about less whimsical examples. Because, you know, theres a lot of researchers and practitioners who are doing the same things. And I guess the concern that I had is: I can tell this output is silly, but can they? And how are they able to vet their own results in kind of the same way that we can when we have a hilarious recipe?

Shane: Yeah, that is the thing about these kinds of experiments that kind of led me into explaining AI. When you have these absurdist, real-world examples Its not deep into the statistics of handing out loans or, you know, looking at resumes, but its messing up something that we all can kind of see. That is a really helpful way of seeing what kinds of limitations there are and how these same sorts of limitations pop up over and over again.

Theyre making these same sorts of mistakes, and the question is, are they looking at them closely enough to catch these mistakes?

VentureBeat: Are there established solutions for people to vet that output, or is that still a big problem to solve?

Shane: The only general rule I can say is, never trust a neural net. Never trust it to know what you asked it. Never trust it not to take a speedy shortcut. And then there are ways of digging in and finding out whether it has solved this correctly. That varies depending on what problem youre trying to solve.

So, in image recognition, there are some really nice tools for explainability, where [the neural net] highlights parts of the images its using to make decisions and [you can] say, Oh no, its supposed to be looking at a dog, and instead its looking at the grassy background, or something like that.

There are those kinds of tools. Some feel explainability is really hard to build in, and in some cases your main tools may be running statistics on your output and saying, Okay, we dont know how its making decisions on these loan applications. We dont know what rules its applying, but we can at least run a whole bunch of test applications. Throw them at it. See if there are any trends in the way that it makes the decision, so we know how the decision should turn out if the thing isnt biased.

I know there are services out there now systematically testing to see whether ones algorithm is biased or has some other kinds of outcomes. And I know that theres a lot of companies that just plain dont do that. You know, I would like to see there be more of a top-down requirement that these kinds of algorithms be vetted by some standardized process at least demonstrate that theyre not problematic. Or theyre not as problematic as they could be.

VentureBeat: That kind of speaks a bit to what you got into in the last couple of chapters, especially when you kind of hammered on human-faked AI: Sometimes these [AI] companies are like, We can do this, trust us, and then it turns out they cant do it, and then they kind of have to panic and either give back all the startup money because they failed or they get in there and kind of mess with it and fix it by hand. That speaks to a bit of the black-boxness of it.

Im coming from a journalist perspective, where Im not an expert. And so we have to have, you know, a certain amount of trust when someone tells us their thing works its tricky. So Im wondering, just in your opinion, how pervasive do you think that problem is in the market?

Shane: I think its pretty pervasive, actually. There are some fields that are perhaps more prone than others to these kinds of problems. I would look really closely at anybody who automatically sorts resumes. Weve seen some case studies, like this one from Amazon, where they volunteered that they had a tool they decided not to use because it was so hard to get it to stop discriminating.

Its kind of like a rare case where we get to see the struggles of the engineering team, trying to get rid of this things tendency to try to copy the bias of training. For every Amazon that tells us, Whoops, we had a problem, we decided to mitigate it by in this case not using this kind of solution at all, weve got a bunch of companies out there who are doing basically the same kind of application, and essentially the same kind of approach, and they havent published any numbers on what kind of bias they have or what specifically theyve done to reduce bias. So, yeah, theres a lot of unfounded claims out there you say, Oh yes, use our computer algorithm and we will take care of your HR teams hiring bias, without any proof that using AI is actually going to help.

VentureBeat: Pretty quickly a trend has emerged, where [makers of AI products and services] are like, Oh no, theres humans in the loop for sure. And whether its lip service or if they really believe it, theyll say, Yeah, look, we have to have the human touch, and they dont come out and say We cant trust our models. With that in mind, do you think that sometimes companies are kind of just pivoting to that messaging without really addressing the core problem?

Shane: There are definitely a lot of AI applications where human curation is needed as a kind of quality control. Definitely in all the AI-generated humor that Im doing, like most of its not interesting. You need a human. I would never train a text-generating algorithm and then just, like, give it to children to have fun with it. It is so hard to sanitize! When people talk to me about doing this kind of project, I always say, You dont want this thing talking to the public. Because it will say something terrible, and you need the human layer in between to stop that from happening.

And, you know, in so many cases, especially with some kind of creative application those are the ones I know the most about I definitely see, you know, its a time saver, but you still need a human. Language translation [is the] same sort of thing; human translators now use machine translation for the first draft. And its close, but it is definitely not ready without some human quality cleanup.

But then we have this other case, going back to having a human in the loop to try to prevent the algorithm from being biased. And thats kind of interesting; circling back to [the idea that] the humans [who built the algorithm] are biased, the algorithms biased and needs the humans. And to that I would just say, Well just, you know show me the data.

We can run test data through this thing. Thats the beauty of having an algorithm without a human running these decisions. We can give it 10,000 loans or 10,000 resumes or 10,000 social media profiles and see if there are any trends. And If someone hasnt done that, I worry that theyre not serious about whether or not their algorithm is flawed in a potentially harmful way.

VentureBeat: What do you think, if anything, has changed in the field like in terms of research, impact, deployment, whatever since you finished writing the book? (I know it was 2019, which is recent.)

Shane: Oh man, things change so quickly. One of the things Ive been encouraged to see is more pushback against some of the bad applications and more communities and people stepping up to bat against this and governments also trying to step in especially [the] European Parliament trying to step in and do the right things.

So Im encouraged that were hopefully going to be a bit out of the snake-oil realm. Therere now more applications out there to worry about, like with facial recognition. Its not great, but its working better, so there are different aspects to worry about, versus in my book, where the concern was [that] it doesnt even work. Well now it kind of works, and thats also a problem.

VentureBeat: What are your top five favorite random things generated by a neural net?

Shane: Oh man. [pause] I really like the experiments where people take the neural net stuff at the starting point and just run with it. There was an experiment I did AInktober where it was generating drawing boxes during the month of October. People drew the most amazing things.

There was one called SkyKnit, where its generating knitting patterns, and people had to try and debug all of the mistakes and make them into things, and it was just glorious.

And then Hat 3000 did the same thing with crocheting. Turns out, using crocheted hats was an unwittingly bad move that tended to create universe-devouring, hyperbolic yarn monstrosities.

There was one example I did where I was generating cookies, and people actually made recipes based on the illustrations, like spice biggers and walps and apricot dream moles.

The paint colors keep coming back again. Using paint colors gives me the opportunity to print the word TURDLY in giant letters across presentation screens.

Signed copies of the book are available from Boulder Book Store.

Read more:

Janelle Shane explains AI with weirdness and humor, in book form - VentureBeat