Artificial Intelligence – an overview | ScienceDirect Topics

Machine learning tools in computational pathology: types of artificial intelligence

AI is not really a new concept. The term AI was first used by John McCarthy in 1955 [4]. He subsequently organized the Dartmouth conference in 1956 which started AI as a field. The label AI means very different things to different observers. For example, some commenters recognize divisions of AI as statistical modeling (calculating regression models and histograms) versus machine learning (Bayes, random forests, support vector machines [SVMs], shallow neural networks, or artificial neural network) versus deep learning (deep neural networks and CNNs). Others recognize categories of traditional AI versus data-driven deep learning AI. In this comparison, traditional AI starts with a human understanding of a domain and seeks to condition that knowledge into models which represent the world of that knowledge domain. When current lay commentators refer to AI, however, they are usually referring to data-driven deep learning AI, which removes the domain knowledge inspired feature extraction part from the pipeline and develops knowledge of a domain by observing large numbers of examples from that domain.

The design approaches of traditional AI versus data-driven deep learning AI are quite different. The architects of traditional AI learning systems focus on building generic models. They often begin with a human understanding of the world through the statement of a prior understanding of the domain (see Fig.11.1), develop metrics representing that prior, extract data using those metrics, and ask humans to apply class labels of interest to these data. These labels are then used to train the system to learn a hyperplane which separates one class from another. Traditional AI learning systems will often be ineffective in capturing the granular details of a problem, and if those details are important, a traditional AI learning system may model poorly.

A data-driven deep learning machine learning system on the other hand can capitalize on the capture of fine details of a system, but it may not illuminate an understanding of the big picture of the problem. Data-driven models are sometimes characterized as black box learning systems which produce classifications or transformed representations of real-world data but without an explanation of the factors that influence the decisions of the learning system. Traditional and deep learning models are compared in Table11.1.

Data-driven deep learning AI approaches have limited humanmachine interactions constrained to a short training period from human annotated data and human verification of the classifier output of the learning system. In contrast, in traditional AI learning systems, human experts can provide actionable insights and bring these rich understandings to the learning system in the form of a prior understanding of the domain. A prior can function as an advanced starting point for a deep learning AI system. The broad understanding of the world that humans possess with their reasoning and inferencing abilities, efficiency in learning, and the ability to transfer knowledge gained from one context to other domains is not very well understood. Framing data-driven deep learning systems with the human understanding of what is offers a way forward for creating partnerships between HI and AI in advanced learning systems. There is need for explainable AI (XAI) which can explain the inferences, conclusions, and decision processes of learning systems. There is much work that needs to be done to bridge the gap between machine and HI.

See more here:
Artificial Intelligence - an overview | ScienceDirect Topics