Core Projects
The original ULTRATHINK research projects, focused on foundational spiking neural network and rank-order coding techniques.
Project Summary
| Project | Directory | Purpose | Agent | Has Code |
|---|---|---|---|---|
| ROSA ViT | Papers/paper/ | Rank-Order Spiking Attention for vision transformers | rosa-agent | No |
| Grandmother | Papers/paper3/ | Invertible Rank-Order Autoencoder | grandmother-agent | Yes |
| SpikeGPT | Papers/paper4/ | O(N) Attention via Spike Timing for language | spikegpt-agent | No |
| Memory Engine | Papers/paper6/ | Hebbian Memory for Continual Learning | memory-agent | No |
| Imagination | Papers/paper7/ | SNN-based Video Generation | imagination-agent | No |
| Audio Engine | Papers/paper8/ | SNN Audio Processing | audio-agent | No |
ROSA ViT (paper/)
Rank-Order Spiking Attention applied to Vision Transformers. The core technique encodes input features as spike timing patterns (rank-order coding) and computes attention using temporal coincidence detection rather than dot-product similarity.
Grandmother (paper3/)
An invertible autoencoder using rank-order coding, inspired by the "grandmother cell" hypothesis in neuroscience. Features working code implementation.
SpikeGPT (paper4/)
Adapts the GPT architecture to use spike-timing-based attention, achieving O(N) complexity instead of O(N^2) by leveraging temporal sparsity in spike trains.
Memory Engine (paper6/)
Implements Hebbian learning rules for continual learning in spiking networks, enabling memory consolidation without catastrophic forgetting.
Imagination (paper7/)
SNN-based generative model for video, producing temporally coherent frame sequences through spike-based dynamics.
Audio Engine (paper8/)
Applies spiking neural networks to audio processing tasks, leveraging the natural temporal structure of audio signals.