Investigations into Learning Algorithms in Intelligent Machines

Abstract

Alan Turing, a famous mathematician, computer scientist, and a World War II code breaker asked a profound question “Can Machines Think?” in his 1950 paper titled “Computing Machinery and Intelligence” published in the journal Mind. In this work, Turing designed a test to determine a machine’s ability to exhibit intelligent behaviour. This test is popularly known as the Turing Test. This foundational work of Turing has inspired researchers to delve deeper into the notion of intelligence. The term Artificial Intelligence (AI) was first coined by John McCarthy who was the organiser of the 1956 Dartmouth Summer Research Project on Artificial Intelligence. Since then, there have been several methods developed to solve real-world problems intelligently by reducing human interventions. But none of them have succeeded in developing a replica of human intelligence. Today, we realise the enormous complexity involved in defining the term intelligence. At present, the field of AI can be seen as a unifying theme of a diverse set of methods/algorithms. These algorithms can be broadly classified as (1) Symbolic AI (logic based systems to replicate rational thought in humans), (2) Machine Learning (algorithms that learn from data), and (3) Sub-symbolic AI (biology/brain inspired learning). From 2010 onwards, researchers showcased the unreasonable effectiveness of brain inspired learning algorithms like deep neural networks, recurrent neural networks, convolutional neural networks etc. Despite their tremendous applications, there is a wide research gap between Artificial Neural Networks (ANNs) and Biological Neural Networks (BNNs). The neurons in BNNs are intrinsically nonlinear and exhibit a wide variety of firing patterns. There is empirical evidence of Deterministic Chaos (random-like behaviour from a deterministic nonlinear system with sensitive dependence on initial conditions) at different spatiotemporal scales in the brain. On the other hand, neurons in ANNs perform only a simple affine transformation followed by nonlinear activation. Also interestingly, the phenomena of stochastic resonance or noise enhanced signal processing is found in the brain. None of these properties are incorporated in current learning algorithms.