Can associative learning be the general process for intelligent behavior in non-human animals?

Abstract

The general process- and adaptive specialization hypotheses represent two contrasting explanations for understanding intelligence in non-human animals. The general process hypothesis proposes that associative learning underlies all learning, whereas the adaptive specialization hypothesis suggests additional distinct learning processes required for intelligent behavior. Here, we use a selection of experimental paradigms commonly used in comparative cognition to explore these hypotheses. We tested if a novel computational model of associative learning -- A-learning -- could solve the problems presented in these tests. Results show that this formulation of associative learning suffices as a mechanism for general animal intelligence, without the need for adaptive specialization, as long as adequate motor- and perceptual systems are there to support learning. In one of the tests, however, the addition of a short-term trace memory was required for A-learning to solve that particular task. We further provide a case study showcasing the flexibility, and lack thereof, of associative learning, when looking into potential learning of self-control and the development of behavior sequences. From these simulations we conclude that the challenges do not so much involve the complexity of a learning mechanism, but instead lie in the development of motor- and perceptual systems, and internal factors that motivate agents to explore environments with some precision, characteristics of animals that have been fine-tuned by evolution for million of years. Author summaryWhat causes animal intelligence? One hypothesis is that, among vertebrates, intelligence relies upon the same general processes for both memory and learning. A contrasting hypothesis states that important aspects of animal intelligence come from species- and problem specific cognitive adaptations. Here, we use a recently formulated model of associative learning and subject it, through computer simulations, to a battery of tests designed to probe cognitive abilities in animals. Our computer simulations show that this associative learning model can account well for how animals learn these various tasks. We conclude that a major challenge in understanding animal and machine intelligence lies in describing behavior systems. Specifically, how motor flexibility and perceptual systems together with internal factors allow animals and machines to navigate the world. As a consequence of our results, together with current progress in both animal- and machine learning, we cannot reject the idea that associative learning provides a general process for animal intelligence.