Reproduce “Deep Learning for Visual Neuroprosthesis” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Peter Beech, Shanshan Jia, Zhaofei Yu, Jian K. Liu
The chapter surveys how deep-learning models relate visual scenes to neural activity. It explains encoders that predict neural responses and decoders that infer scenes, including Spike Image Decoder, which maps retinal ganglion-cell spikes to an intermediate image before refinement. Prior work reconstructed scenes from isolated animal retinas, demonstrating decoding under experimental conditions, not prosthesis control or restored human vision.
Key insights
- Deep networks can provide useful approximations of visual representations across retinal and cortical areas, but decoding success does not necessarily show that the model has recovered the biological mechanism of perception.
- Generative priors improve the plausibility and semantic content of reconstructions, while also allowing models to insert features inferred from training data rather than explicitly contained in the neural measurements.
- For prosthesis design, the most useful role of deep learning may be to estimate biologically meaningful stimulation patterns or evaluate the quality of prosthetically generated neural activity, rather than simply maximize image similarity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.