Hey all!
I’m startup founder and neuroscience PhD who has been working on a bio-inspired PyTorch plugin. It is based on a novel adjustment to the backpropagation algorithm that allows models to continue improving their accuracy after being fully trained. In some cases this allows you to start with a smaller network and recover your original accuracy while maintaining a significantly reduced parameter count. I currently have 4 results where I have pushed a state-of-the-art system with it, though for transparency it doesn’t work in all settings. I’m looking for collaborators who would like to try it out on their systems to validate the results in additional projects. You will have to run training again, but the actual coding adjustments will only take 5 minutes to an hour depending on the complexity of your network. Please comment or DM me if you’d like to try it out!
Thanks!