Artificial intelligence, by Charu C. Aggarwal, covers the breadth of knowledge in the field of artificial intelligence (AI), from its beginnings to how it took a turn in recent years. The book starts by explaining the two major schools of thought that define current AI as it develops today in different areas such as game playing, self-driving autonomous cars, image processing, expert systems and robotics, and so on. The author has thoroughly researched all areas of AI in order to write this high-quality book.
Aggarwal begins with deductive reasoning, which was a major driving force in the development of AI though it did not lead to much success in the field. Later, inductive learning such as machine learning and deep learning, together with deductive reasoning, brought many achievements in the industry. The successes were not only for the industry, however, but for humanity as a whole. This also slowly brought AI closer to the biological paradigm due to the fact that more and more people incorporated AI in daily life. Massive amounts of data were collected, which led to more computational requirements. Having several cores along with the emergence of graphics processing units (GPUs) have greatly impacted the field. Search algorithms continue to play a key role.
The author goes on to explain various search algorithms, including uninformed depth-first search and breadth-first search. The search algorithm started to improve with the use of state-specific loss values, which led to its incorporation in several applications. Single agent developed into multi-agent and inductive learning methods such as Monte Carlo simulation were incorporated.
In the second part, the author explains propositional logic, which is a mathematical foundation for expert systems. This led to the construction of a first-order logic that helped to develop expert systems and logic-based approaches in AI. As a result, first-order logic was applied in several real-world settings, but for narrow domains only. Since it would require very large knowledge bases and can’t meet the computational complexity, it was not applied to generalized forms of AI.
Human cognition is also based on evidentiary experiences, not just using logic-based methods alone. In the third part of the book, the author presents inductive learning methods in detail. In machine learning, “not all predictions ... can be performed in an interpretable way, simply because many of the choices that humans make ... are not easily expressible in terms of interpretable choices.” Hence, several machine learning methods and various techniques for the evaluation of classification and regression are explained in detail.
The fourth part of the book explores the emergence of neural networks in basic computational graphs for machine learning applications. Learning the parameters of a computational graph from observed data provides a route to learning functions from observed data. Thus, traditional neural networks provide a special case of graphs. This is the case for acyclic graphs. Other types of undirected and cyclic graphs are also used to represent models like Hopfield networks and restricted Boltzmann machines. Further, several domain-specific neural architectures are explained: convolutional neural networks (CNNs) and recurrent neural networks (RNNs), including discrete RNN, word-level RNN, and character-level RNN. Many applications of recurrent and convolutional networks are also explained in detail. The above-mentioned supervised learning accuracy can be further improved by using unsupervised learning. In unsupervised learning, various techniques such as the compressed representation of data (for example, reduced dimensionality, linear dimensionality reduction) are used to represent the data. This is closely related to clustering, which can lead to manifold reduction in the labeled data that is required for classification.
The book ends with reinforcement learning in which many methods “integrate deep neural networks to take in sensory inputs and learn policies that optimize rewards.” Even though these algorithms learn via experimentation, which “leads to innovative solutions,” probabilistic graphical models and knowledge graphs are also used in several applications, including the semantic web, knowledge-based representation, and heterogeneous information network analysis. Finally, the book stresses the need for the integration of inductive learning and deductive reasoning methods, which can often lead to more accurate results even for small datasets.
This highly valuable book provides a vast overview of AI in a well-structured manner. It could be used as a textbook in graduate-level courses.