Interactive
CWM Lab
Run the glass memory experiment on your own resonator — or explore with simulation. Encode messages into vibrational fingerprints, train an Echo State Network, and decode sequences from memory.
Sign in to save your work
Your resonator configuration, calibration data, and experiment results are stored in the cloud. Sign in with Google to persist across sessions and resume calibrations if they're interrupted.
Choose Your Setup
Calibrate & Train
Select a setup above, then calibrate to build the feature map and train the ESN.
Decode a Message
ℹ️ How does this work?
The experiment: An Echo State Network (ESN) — a recurrent neural network with a fixed random reservoir — reads a 4-character sequence and must output it reversed. This requires genuine memory: the network must hold all 4 characters before emitting any output.
Glass path: Each 8-bit ASCII character is encoded by driving a subset of the resonator's eigenmodes. The physical coupling between modes creates a unique spectral fingerprint per token — richer than the binary input because nonlinear anharmonic coupling adds information the software baseline doesn't have. Multiple resonators multiply the feature space.
Software baseline: The same 8-bit binary encoding, expanded to 163 dimensions via degree-4 polynomial interaction terms. Despite having 10× more features, the polynomial expansion overfits with the same training data.
Why glass wins: The physics naturally produces well-conditioned features. The polynomial baseline is overcomplete → ridge regression overfits. Glass features are information-dense because the coupling is governed by the rod's actual mechanical transfer function, not a generic expansion.