just idea..
Two suggestions for an input/output domain beyond MNIST.
Input: biosignals (EEG/EMG). A signal window folds into the same tensor
shape the model already takes — 16 channels x 16 time bins, or 8 channels
x 32 band-power bins = 256 uint8 inputs. No change to the input layer.
Class counts are small (blink, jaw clench, left/right, idle), and public
datasets exist (e.g. PhysioNet EEG Motor Movement/Imagery). EMG is the
easier first target: much higher SNR than motor imagery, good for
validating the pipeline.
Output: USB HID via bit-banged USB (rv003usb). That closes the loop
entirely on the MCU — electrodes in, keystrokes out, no host-side
classifier and no driver. Existing BCI boards (e.g. the recent Octopus 16,
16x 24-bit channels over SPI to a XIAO ESP32-S3) stream raw data to a PC
and classify there; this may be the same thing at a fraction of the
part cost.
ref: https://www.cnx-software.com/2026/08/02/play-games-with-your-brain-signals-using-octopus-16-wireless-eeg-device/
just idea..
Two suggestions for an input/output domain beyond MNIST.
Input: biosignals (EEG/EMG). A signal window folds into the same tensor
shape the model already takes — 16 channels x 16 time bins, or 8 channels
x 32 band-power bins = 256 uint8 inputs. No change to the input layer.
Class counts are small (blink, jaw clench, left/right, idle), and public
datasets exist (e.g. PhysioNet EEG Motor Movement/Imagery). EMG is the
easier first target: much higher SNR than motor imagery, good for
validating the pipeline.
Output: USB HID via bit-banged USB (rv003usb). That closes the loop
entirely on the MCU — electrodes in, keystrokes out, no host-side
classifier and no driver. Existing BCI boards (e.g. the recent Octopus 16,
16x 24-bit channels over SPI to a XIAO ESP32-S3) stream raw data to a PC
and classify there; this may be the same thing at a fraction of the
part cost.
ref: https://www.cnx-software.com/2026/08/02/play-games-with-your-brain-signals-using-octopus-16-wireless-eeg-device/