The author began investigating ways to interpret multi-layer perceptron (MLP) neural networks by generating training data from simple equations. After failing to develop a lossless compression algorithm for Neuralink, the author turned a children’s toy into an EEG to collect data at home for neural network training. The Neurosky EEG chip was used to train a neural network to predict button depression based on brainwave signals. An interpretability engine helped simplify the neural network into singular perceptions, leading to the development of more efficient models with improved execution times. This study showcases the potential for enhancing neural network inference efficiency in Neuralink applications.
https://geofflord.substack.com/p/brainwaves-on-a-budget