This repository contains a solution for the [Neuralink Compression Challenge](https://content.neuralink.com/compression-challenge/README.html). The challenge involves compressing raw electrode recordings from a Neuralink implant. These recordings are taken from the motor cortex of a non-human primate while playing a video game.
The Neuralink N1 implant generates approximately 200 Mbps of electrode data (1024 electrodes @ 20 kHz, 10-bit resolution) and can transmit data wirelessly at about 1 Mbps. This means a compression ratio of over 200x is required. The compression must run in real-time (<1ms)andconsumelowpower(<10mW,includingradio).
The `analysis.ipynb` notebook contains a detailed analysis of the data. We found that there is sometimes significant cross-correlation between the different leads, so we find it vital to use this information for better compression. This cross-correlation allows us to improve the accuracy of our predictions and reduce the overall amount of data that needs to be transmitted. As part of the analysis, we also note that achieving a 200x compression ratio is highly unlikely to be possible and is also nonsensical; a very close reproduction is sufficient.
As the first step, we analyze readings from the leads to construct an approximate topology of the threads in the brain. The distance metric we generate only approximately represents true Euclidean distances, but rather the 'distance' in common activity. This topology must only be computed once for a given implant and may be updated for thread movements but is not part of the regular compression/decompression process.
The main workhorse of our compression approach is a predictive model running both in the compressor and decompressor. With good predictions of the data, only the error between the prediction and actual data must be transmitted. We make use of the previously constructed topology to allow the predictive model's latent to represent the activity of brain regions based on the reading of the threads instead of just for threads themselves.
1.**Latent Projector**: This module takes in a segment of a lead and projects it into a latent space. The latent projector can be configured as a fully connected network or an RNN (LSTM) with an arbitrary shape.
2.**MiddleOut (Message Passer)**: For each lead, this module performs message passing according to the thread topology. Their latent representations along with their distance metrics are used to generate region latent representation. This is done by training a fully connected layer to map from (our_latent, their_latent, metric) -> region_latent and then averaging over all region_latent values to get the final representation.
3.**Predictor**: This module takes the new latent representation from the MiddleOut module and predicts the next timestep. The goal is to minimize the prediction error during training. It can be configured to be an FCNN of arbitrary shape.
The neural networks used in this solution are rather small, making it possible to meet the latency and power requirements if implemented more efficiently.
- All currently implemented bitstream encoders are rather naive. We know that lead values from the N1 only have 10-bit precision, but the WAV file provides us with 32-bit floats. All my bitstream encoders are also based on 32-bit floats; discretizing back into the 10-bit space would be a low-hanging fruit for ~3.2x compression.
- Since we merely encode the remaining delta, we can go even more efficient by constructing something along the lines of a Huffman tree.
- Loss is not coming down during training... So basically nothing works right now. But the text I wrote is cool, right?
