Building an EEG with a Children's Toy

In February 2024, Geoff Lord began exploring the interpretability of multi-layer perceptron (MLP) neural networks, aiming to generate training data from a simple equation and rediscover it through model parameters. After unsuccessful attempts with the Neuralink Compression Challenge, he creatively repurposed a Mattel Mindflex toy to collect EEG data. The toy's EEG chip was modified to access neural activity data, allowing for the collection of samples while alternating between relaxed and attentive states. The data was used to train an MLP neural network to predict button presses based on EEG signals, achieving 68.68% training accuracy and 58.00% testing accuracy. The study highlighted the importance of neural network interpretability, revealing that the high alpha waveform was the most correlated input for predicting button presses. This insight led to the development of optimized models that significantly reduced execution time—up to 95.18%—while maintaining comparable accuracy. The findings suggest that interpretability and optimization techniques can enhance neural network efficiency, potentially benefiting applications in Neuralink and similar technologies.

- Geoff Lord repurposed a children's toy to collect EEG data for neural network training.

- The study achieved notable improvements in model execution time and accuracy through interpretability techniques.

- High alpha waveform was identified as the most significant input for predicting actions.

- Optimized models demonstrated execution time reductions of up to 95.18%.

- The research contributes to advancements in neural network inference efficiency for applications like Neuralink.