Please note: This PhD defence will take place in DC 2314 and online.
Ehsan Ganjidoost, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Jeff Orchard
Standard deep learning architectures trained via gradient backpropagation diverge substantially from biological neural processing due to non-local weight transport, update locking, and rigid global error computation. Furthermore, feedforward discriminative models remain susceptible to imperceptible adversarial perturbations, struggle with multi-cluster and multimodal distributions, and lack the recurrent generative feedback characteristic of cortical circuits. Grounded in Karl Friston’s Free Energy Principle, the predictive coding network (PCnet) offers a biologically plausible alternative by performing continuous, bidirectional message passing where internal representations and localized Hebbian synaptic updates minimize hierarchical prediction errors.
This dissertation investigates the theoretical mechanics, computational scaling, and practical architectures of PCnets across associative memory, generative multimodal modelling, and adversarial defence. First, we establish the dynamical foundations of PCnets, showing that continuous energy minimization yields exact Maximum A Posteriori inference and Lyapunov stability under local plasticity. At the same time, semi-dynamic approximations reduce training iterations without destabilizing parameter equilibria.
Second, we reformulate associative memory retrieval through hierarchical error minimization by introducing a predictive coding Hopfield network (PCHN) and its stochastic variant, stochastic PCHN (sPCHN). Compared with the original (classic) Hopfield network (OHN) and its stochastic counterpart, stochastic OHN (sOHN), our models eliminate prototype collapse and recall bias. Furthermore, by adding controlled stochastic noise during recall, the network easily escapes ambiguous states and spurious local traps, reliably restoring corrupted patterns and identifying local cluster centres in complex data distributions.
Third, to mitigate mode collapse and improve multimodal learning in deep networks, we introduce associative predictive coding (PC) architectures. By connecting the internal representations of separate sensory pathways, the network captures complex variations without sacrificing classification performance. This shared representation enables the model to infer missing inputs and generate consistent data across different modalities.
Finally, this thesis proposes an adversarial defence mechanism utilizing the recurrent generative feedback of (PCnet). Deployed as a model-agnostic pre-inference filter for downstream feedforward classifiers, the architecture interprets imperceptible, human-undetectable adversarial perturbations as prediction error. It dynamically purifies corrupted activations back toward clean data manifolds before final classification.
Together, these contributions demonstrate that hierarchical predictive coding provides a unified, computationally efficient, and biologically grounded foundation for scalable memory systems, multi-sensory integration, and robust neural computation.
To attend this PhD defence in person, please go to DC 2314. You can also attend virtually on MS Teams.