Generalization of information - integrative encoding or category-based inference?

Jessica Taylor, Helen Barron, Dasa Zeithamova, Masamichi Sakagami, Aurelio Cortese, Xiaochuan Pan

• Abstract:

  Title: Generalization of information - Integrative encoding or category-based inference?
  Scientific question: How do biological organisms generalize previously-learned information for adaptive behaviour in novel experiences? This broad question spans interdisciplinary fields - from decision-making and perception, psychology, memory neuroscience, to theoretical considerations in machine learning and artificial intelligence (Cortese et al. 2019). One prominent theory (integrative encoding) suggests that when new memories are being made, we activate information from overlapping memory representations so that information from both are integrated and reencoded together. Behaviour in novel situations can be guided by simple recall of information related to these extended associative links. Although this mechanism is very simple, computationally this process could become tedious given the inestimable extent of potential (and sometimes unnecessary) associative links that could be formed. A different theory (category-based inference) suggests a computationally simpler mechanism: that humans use abstract thoughts, such as functional categorization, to make their behaviour more efficient. Categories provide a logical structure through which information learned for one stimulus may be generalized to other stimuli (members of the same category). However, this theory requires much higher order, complicated, abstractions than does the former. One possibility is that both of these theories might be valid, with an organism implementing different strategies dependent on their current circumstances. However, to date there exists no study that has clearly dissected the contribution of both theories, nor made formal predictions on how they could or should differ in neural or behavioural terms.