Recall a sequence.

Two targets hide among six distractors. When the two prompts arrive, a four-cell LSTM must reproduce the targets in their original order. Watch the network receive and answer one symbol at a time.

training WebGPU
targets to remember distractors recall prompts no output expected

Live sequence

waiting for training
Training a length 4→10 curriculum.
The sequence player will unlock when the model is ready.
Trial0 / 9,625
Curriculumlength 4
Train time
Last sequence

Building recurrent memory through progressively longer sequences…