Representation learning with generative neural networks A biological perspective

This thesis investigates the computational mechanisms underlying generative representation learning in biological perception and explores biologically plausible extensions to predictive coding (PC) models. We address four key research questions concerning representation learning, disentanglement of sensory input causes, and alignment with neural dynamics, and variations in the generative task beyond classical predictive coding.
Chapter 2 demonstrates that temporal statistics can serve as an inductive bias for PC networks, enabling invariance to input transformations. Training on continuous input sequences fosters a hierarchical timescale structure in network representations, consistent with neural dynamics observed in the ventral visual stream. The trained networks are capable of generative image reconstruction, even under occlusion.
Chapter 3 introduces a novel PC architecture that disentangles optic flow generated by self-motion from externally caused patterns. The biologically plausible model incorporates sensorimotor mismatch circuits and accurately segments moving objects from backgrounds, aligning with experimental calcium imaging data in mice.
In Chapter 4, we propose a self-supervised, generative task inspired by eye movements and primate foveal vision. A proof-of-concept model using masked image modeling demonstrates that peripheral masking decorrelates latent space neurons and improves downstream classification compared to autoencoding full visual inputs.
Together, these findings highlight the potential of generative neural networks to solve image processing tasks and to account for information processing in the mammalian visual cortex.