Brain-like Representation Learning and Associative Memory

QC 20260907

Abstract

The brain enables organisms to perceive the world, learn from experience, generate complex behavior, and give rise to cognition. Understanding the information processing principles underlying brain computation remains a key challenge in computational neuroscience and cognitive science. Elucidating these principles has the potential to provide a foundation for developing intelligent machines and energy-efficient, scalable, and robust artificial intelligence paradigms.

This thesis investigates neural network models incorporating key brain-like design principles, focusing on unsupervised representation learning and the formation of robust associative memory. While modern deep learning approaches achieve strong performance in representation learning tasks, they rely on backpropagation-based optimization, which lacks biological plausibility and depends on globally coordinated computations. There remains a need for neural network models that can demonstrate complex functionalities while remaining grounded in biological principles.

To address this gap, this thesis develops a class of brain-like models based on the Bayesian Confidence Propagation Neural Network (BCPNN) framework. The proposed models incorporate key brain-like design principles, including localized Hebbian synaptic plasticity, structural plasticity, activity and connection sparsity, and a modular architecture derived from neocortical columnar organization. Furthermore, the models integrate feedforward, recurrent, and feedback connectivity to support both representation learning and associative memory.

This work comprises three main lines of investigation. First, the unsupervised representation learning capabilities of a feedforward model are examined, demonstrating that structured internal representations can be learned directly from unlabeled data using localized synaptic and structural plasticity. Second, the formation of associative memory is studied by integrating recurrent connectivity in the representation learning model, demonstrating robust recall from partial, noisy, or corrupted inputs when tested on pattern completion, prototype extraction, and noise robustness tasks. Third, the framework is extended to spiking neural networks, where neurons communicate via stochastic spike events, and the results show that, through synaptic short-term filtering and appropriate temporal scaling, the spiking models approximate the behavior of their rate-based counterparts while preserving functionality and performance. 

Overall, the results establish that brain-like design principles can support scalable representation learning and robust associative memory, demonstrating a bridge between neuroscience and artificial intelligence within the emerging field of NeuroAI. This work further offers a pathway toward energy-efficient, brain-like neuromorphic systems capable of operating in real-time and in dynamic environments.

Link to DiVA