I'm building an open-source cognitive agent (Bride) that combines HDC for semantic memory with Active Inference for decision-making. Architecture: ~10k-dim hypervectors, Ebbinghaus forgetting, Free Energy minimization.
Three practical questions for the HDC community:
1. Vector similarity drift: We use cosine similarity between predicted and observed HDC vectors as the observation signal for AIF. Over time, similarity distributions compress toward 0.5-0.7 (all concepts start looking the same). Any known techniques to maintain vector diversity/discriminability?
2. Bundling for sequential memory: When storing temporal sequences (A→B→C), we bundle successive vectors with a decay factor. Works for ~5 steps before the sequence is unrecoverable. Is there a better approach using binding + permutation?
3. Hardware acceleration: torchhd works great on CPU but our 10k-dim operations are the bottleneck. Any recommendations for GPU/MPS optimization patterns?
Project: https://github.com/cedendahlkim/bride-cli | API: https://tools.gracestack.se/oracle-bridge-landing.html
Would appreciate any pointers from folks working at this intersection.
I'm building an open-source cognitive agent (Bride) that combines HDC for semantic memory with Active Inference for decision-making. Architecture: ~10k-dim hypervectors, Ebbinghaus forgetting, Free Energy minimization.
Three practical questions for the HDC community:
1. Vector similarity drift: We use cosine similarity between predicted and observed HDC vectors as the observation signal for AIF. Over time, similarity distributions compress toward 0.5-0.7 (all concepts start looking the same). Any known techniques to maintain vector diversity/discriminability?
2. Bundling for sequential memory: When storing temporal sequences (A→B→C), we bundle successive vectors with a decay factor. Works for ~5 steps before the sequence is unrecoverable. Is there a better approach using binding + permutation?
3. Hardware acceleration: torchhd works great on CPU but our 10k-dim operations are the bottleneck. Any recommendations for GPU/MPS optimization patterns?
Project: https://github.com/cedendahlkim/bride-cli | API: https://tools.gracestack.se/oracle-bridge-landing.html
Would appreciate any pointers from folks working at this intersection.