Ananya Passi¹, Michael F Bonner¹; ¹Johns Hopkins University

Deep neural networks (DNNs) are leading computational models of the visual cortex. However, their reliance on biologically implausible training methods, such as task supervision and backpropagation, limits their interpretability and alignment with biological principles. Additionally, the time- and resource-intensive nature of their training discourages iterative exploration, further hindering insights into high-performing models of the visual cortex. In contrast, biological visual representations are believed to emerge through unsupervised, local learning mechanisms that do not rely on backpropagation, highlighting the need for alternative computational approaches that better reflect these processes. Here, we propose a biologically inspired framework for learning hierarchical visual representations using local unsupervised learning without backpropagation. In this approach, each layer of a DNN incorporates a bottleneck mechanism that compresses and subsequently expands representations. The learning process is entirely unsupervised, with each layer optimizing only to compress its inputs. This minimalist algorithm produces representations that rival models trained using backpropagation in their ability to predict image-evoked fMRI responses in the visual cortex up to intermediate processing stages. By aligning with key principles of biological learning, our approach requires no labeled data or task-specific supervision and provides a parsimonious framework for modeling visual hierarchy formation. Our model offers neuroscientists a novel tool for conducting in-silico analyses and controlled rearing experiments with reduced computational overhead. Furthermore, its simplicity and biological plausibility provide new insights into how visual computations might be organized in the brain, advancing our understanding of the neural mechanisms underlying vision.