Avery HW Ryoo
banner
averyryoo.bsky.social
Avery HW Ryoo
@averyryoo.bsky.social
i like generative models, science, and Toronto sports teams

phd @ mila/udem, prev. @ uwaterloo

averyryoo.github.io 🇨🇦🇰🇷
🐐
July 12, 2025 at 8:23 PM
Finally, we show POSSM's performance on speech decoding - a long context task that can quickly grow expensive for Transformers. In the unidirectional setting, POSSM beats the GRU baseline, achieving a phoneme error rate (PER) of 27.3 while having more robustness to variation in preprocessing.

🧵6/7
June 6, 2025 at 5:40 PM
Cross-species transfer! 🐵➡️🧑

Excitingly, we find that POSSM pretrained solely on monkey reaching data achieves SOTA performance when decoding imagined handwriting in human subjects! This shows the potential of leveraging NHP data to bootstrap human BCI decoding in low-data clinical settings.

🧵5/7
June 6, 2025 at 5:40 PM
By pretraining on 140 monkey reaching sessions, POSSM effectively transfers to new subjects and tasks, matching or outperforming several baselines (e.g., GRU, POYO, Mamba) across sessions.

✅ High R² across the board
✅ 9× faster inference than Transformers
✅ <5ms latency per prediction

🧵4/7
June 6, 2025 at 5:40 PM
POSSM combines the real-time inference of an RNN with the tokenization, pretraining, and finetuning abilities of a Transformer!

Using POYO-style tokenization, we encode spikes in 50ms windows and stream them to a recurrent model (e.g., Mamba, GRU) for fast, frequent predictions over time.

🧵3/7
June 6, 2025 at 5:40 PM
The problem with existing decoders?

😔 RNNs offer efficient, causal inference, but rely on rigid, binned input formats - limiting generalization to new neurons or sessions.

😔 Transformers enable generalization via tokenization, but have high computational costs due to the attention mechanism.

🧵2/7
June 6, 2025 at 5:40 PM
@oliviercodol.bsky.social my opportunity to lose to scientists in a different field
March 27, 2025 at 2:06 AM