Grant: Discovery Grants Program - Individual

Across a variety of tasks, a hallmark of learning is that improvements in performance are accompanied by an increase in the predictability of responses. To understand this phenomenon, learning can be split into two components: (1) a "bottom-up" process focused on attaining a desired performance; and (2) a "top-down" process where cognitive expectations drive an increase in predictability. While computational models termed artificial recurrent networks have successfully captured brain activity across several tasks, these models have focused on a bottom-up approach where synaptic weights are tuned based on a desired outcome, while neglecting top-down expectations. The long-term goal of my research program is to provide a comprehensive description of learning mechanisms in recurrent circuits of the brain and accelerate the development of algorithms that achieve human-like performance. In this application, I hypothesize that recurrent networks that incorporate top-down expectations will outperform networks trained with standalone bottom-up learning. To test this hypothesis, I will train recurrent networks using a novel rule that combines bottom-up (supervised) and top-down (predictive) learning. Three specific aims will be pursued: 1. Optimal trade-off between supervised and predictive learning. Recurrent networks will be trained on categorization and time series prediction tasks. In these networks, the parameterization of supervised and predictive learning will allow us to investigate their relative influence. 2. Contribution of localized error signals. In the brain, specialized populations of cells track either positive or negative prediction errors. This observation will be incorporated in artificial networks, resulting in localized error signals that may improve learning accuracy by acting as a form of regularization. 3. Robustness to noisy or mislabelled data. We predict that top-down information will provide recurrent networks with improved robustness, given that it can be shown analytically that predictive learning dampens the magnitude of synaptic modification in the presence of noise. These aims will reveal the advantages of incorporating top-down information in artificial recurrent networks, in terms of speeding up learning times and providing robustness to noisy data. This work will pioneer the emergence of plasticity rules that combine bottom-up and top-down learning, as a gateway to understanding human learning and developing artificial algorithms. Results will be disseminated in peer-reviewed publications aimed at advancing knowledge in fundamental and applied aspects of computational intelligence. Projects based on the three specific aims will contribute to the training of highly qualified personnel at the undergraduate and graduate levels who will develop skills in the areas of computational neuroscience, programming, and formal mathematical analysis of dynamical systems.