Abstract
In recent years, deep learning approaches have experienced tremendous growth in their learning capabilities. These successes have predominantly been driven by new network architectures, larger datasets and more powerful hardware, while the fundamental principles behind the learning algorithms have remained the same. Since its inception, the error backpropagation (backprop) algorithm has been a foundational component for performing credit assignment within artificial neural networks. In the brain, however, such a ubiquitous learning mechanism has yet to be identified, and many have long believed that backprop is biologically implausible or insufficient for explaining the brain's wide array of learning capabilities. Recent models have challenged this notion by demonstrating efficient and biologically plausible backprop-like learning.We introduce a computational framework for the cholinergic neuromodulatory system which is integrated with a cortical network model of biologically plausible backprop. In this framework, the cholinergic system accumulates local error signals received from within the cortical network during task learning and releases neuromodulatory signals that act to modulate the learning rates of that same network. Its function is inspired by adaptive optimisation methods commonly used for training artificial neural networks. However, unlike these methods, which typically produce many independent modulatory signals at the level of individual synapses, this framework produces more diffuse signals which are more in line with observations of the cholinergic system. Our results using this framework show that these diffuse modulatory signals are sufficient to improve learning performance on standard image classification tasks.
Next, we introduce the Bursting Cortico-Cortical Network (BurstCCN), which provides a biologically plausible mechanism of backprop in which error signals are represented in the apical dendrites of pyramidal neurons that control the relative frequency of burst activity. These bursts induce local burst-dependent plasticity and give rise to a burst multiplexing code that facilitates the backwards propagation of errors though connection-type-specific STP feedback connections. We show that the BurstCCN is able to use a single-phase learning process to effectively backpropagate errors signals that both empirically and analytically approximate backprop-derived gradients. We demonstrate the BurstCCN's ability to effectively learn complex image classification tasks (MNIST and CIFAR-10) and from scalar reward signals. Furthermore, we extend the BurstCCN model to enforce Dale's principle, which requires a strict separation of excitatory and inhibitory neuron populations, and show this model still learns successfully.
Overall, this thesis explores many biologically plausible implementations of the cholinergic system and cortical networks, drawing inspiration from deep learning principles.
| Date of Award | 1 Oct 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Rui Ponte Costa (Supervisor) & Jack R Mellor (Supervisor) |