Adaptive Early Vision Networks: Learning Biologically Constrained Front-Ends for Robust Visual Recognition

Deep neural networks have achieved remarkable performance in visual recognition, but they remain more fragile than biological vision systems when images are corrupted, perturbed, or shifted away from the training distribution. In contrast, primate vision is highly robust across changes in contrast, noise, blur, illumination, and viewpoint. This has motivated a growing research area investigating whether principles from neuroscience can help design more robust and interpretable artificial vision systems.

Previous work from our group and collaborators introduced VOneNets, hybrid neural networks that place a biologically inspired model of the primate primary visual cortex (V1) at the front of standard convolutional neural networks. This architecture showed that constraining the early stages of artificial vision using known properties of biological visual processing can improve robustness to several image perturbations. More recently, Early Vision Networks (EVNets) extended this idea by explicitly modeling pre-cortical visual processing stages, including retina- and LGN-inspired computations, before the V1-like stage.

The goal of this Master's thesis is to develop a new generation of biologically inspired neural network architectures in which the early visual front-end is no longer fully fixed, but instead learnable under biological constraints. The project will investigate whether selected parameters of the VOneNet/EVNet front-end — such as spatial frequency tuning, orientation selectivity, center-surround interactions, contrast normalization, or neural noise — can be optimized during training while remaining close to biologically plausible distributions. This would allow the model to adapt to the task and dataset while preserving the interpretability and robustness benefits of biologically inspired design.

The student will implement and evaluate adaptive variants of VOneNet/EVNet architectures, comparing fixed, unconstrained learnable, and biologically constrained learnable front-ends. The project will involve model implementation in PyTorch, training and evaluation on standard computer vision datasets, and systematic ablation studies to understand which biological constraints contribute most to accuracy, robustness, and representation quality.

Possible research directions include making selected V1 or subcortical parameters learnable with regularization toward empirical biological distributions; introducing differentiable constraints on Gabor filters, Difference-of-Gaussian filters, center-surround mechanisms, and normalization parameters; comparing fixed versus adaptive early visual front-ends; evaluating whether biological constraints improve robustness compared with fully learnable alternatives; and exploring whether the adaptive front-end can be combined with different downstream architectures such as ResNets, EfficientNets, ConvNeXt models, or Vision Transformers.