RESUMO
Human ability to recognize complex visual patterns arises through transformations performed by successive areas in the ventral visual cortex. Deep neural networks trained end-to-end for object recognition approach human capabilities, and offer the best descriptions to date of neural responses in the late stages of the hierarchy. But these networks provide a poor account of the early stages, compared to traditional hand-engineered models, or models optimized for coding efficiency or prediction. Moreover, the gradient backpropagation used in end-to-end learning is generally considered to be biologically implausible. Here, we overcome both of these limitations by developing a bottom-up self-supervised training methodology that operates independently on successive layers. Specifically, we maximize feature similarity between pairs of locally-deformed natural image patches, while decorrelating features across patches sampled from other images. Crucially, the deformation amplitudes are adjusted proportionally to receptive field sizes in each layer, thus matching the task complexity to the capacity at each stage of processing. In comparison with architecture-matched versions of previous models, we demonstrate that our layerwise complexity-matched learning (LCL) formulation produces a two-stage model (LCL-V2) that is better aligned with selectivity properties and neural activity in primate area V2. We demonstrate that the complexity-matched learning paradigm is responsible for much of the emergence of the improved biological alignment. Finally, when the two-stage model is used as a fixed front-end for a deep network trained to perform object recognition, the resultant model (LCL-V2Net) is significantly better than standard end-to-end self-supervised, supervised, and adversarially-trained models in terms of generalization to out-of-distribution tasks and alignment with human behavior.
RESUMO
OBJECTIVE: The nature of artificial vision with a retinal prosthesis, and the degree to which the brain can adapt to the unnatural input from such a device, are poorly understood. Therefore, the development of current and future devices may be aided by theory and simulations that help to infer and understand what prosthesis patients see. APPROACH: A biologically-informed, extensible computational framework is presented here to predict visual perception and the potential effect of learning with a subretinal prosthesis. The framework relies on optimal linear reconstruction of the stimulus from retinal responses to infer the visual information available to the patient. A simulation of the physiological optics of the eye and light responses of the major retinal neurons was used to calculate the optimal linear transformation for reconstructing natural images from retinal activity. The result was then used to reconstruct the visual stimulus during the artificial activation expected from a subretinal prosthesis in a degenerated retina, as a proxy for inferred visual perception. MAIN RESULTS: Several simple observations reveal the potential utility of such a simulation framework. The inferred perception obtained with prosthesis activation was substantially degraded compared to the inferred perception obtained with normal retinal responses, as expected given the limited resolution and lack of cell type specificity of the prosthesis. Consistent with clinical findings and the importance of cell type specificity, reconstruction using only ON cells, and not OFF cells, was substantially more accurate. Finally, when reconstruction was re-optimized for prosthesis stimulation, simulating the greatest potential for learning by the patient, the accuracy of inferred perception was much closer to that of healthy vision. SIGNIFICANCE: The reconstruction approach thus provides a more complete method for exploring the potential for treating blindness with retinal prostheses than has been available previously. It may also be useful for interpreting patient data in clinical trials, and for improving prosthesis design.
