On-chip data compression

On-chip data compression reduces the neural data rate before it reaches the wireless link, which is the true bottleneck for high-channel-count implants. Lossy methods dominate because neural data is sparse and structured: detecting and sorting spikes to send timestamps and labels, binning spike-band power, transmitting only thresholded snippets, transform or wavelet coding, and compressive sensing that projects sparse spike activity onto a small number of random measurements. Lossless methods — entropy coding and predictive or delta coding of the residual — give more modest ratios but preserve the waveform.

Typical compression ratios span roughly ten to a hundred times depending on how much information is discarded. The right operating point depends entirely on the downstream task: closed-loop decoding may need only spike-band power at low rate, whereas any hope of offline re-analysis argues for keeping waveforms.

Two constraints keep compression honest. First, it is only worthwhile if the compute power it costs on the implant is less than the transmit power it saves — compression that heats the tissue more than the radio would is counterproductive. Second, aggressive lossy compression is irreversible, so it silently forecloses analyses that were not anticipated at design time.