PCA drops space, HMMs drop time: why existing brain-activity methods lose information
Pseudomanifold · x · 2026-11-09
Post 4/n in the Krishnaswamy Lab thread argues that existing neural-activity analysis methods sacrifice one of two axes.
- Spatial reduction: PCA, ICA, and correlation-based connectivity turn brain states into vectors with no particular spatial organization.
- Temporal reduction: HMMs and dynamic functional connectivity compress dynamics into an average, a single graph, or discrete states.
As the figure stresses, neural activity is fundamentally a spatial pattern at each timepoint evolving over time — reducing either axis causes a large loss of information. This motivates the thread's graph-based approach, which represents activity as a time-varying signal on a graph.
Related event: Modeling Neural Activity as Time-Varying Signals on Graphs(3 posts)→
More from Research
- New theory shows prover-conjecturer self-play can grow proved theorems exponentially — StanfordAILab · 2026-11-09
- AbridgeKV: latent KV cache compression matches full accuracy at 1/4 cache, 3.2× faster — DimitrisPapail · 2026-11-09
- Stanford's Hinted Scaling Laws cut LLM capability forecast error from 21% to 4.3% — burny_tech · 2026-11-09
- Stanford's Evo AI designs 16 working bacteria-killing viruses, sparking access debate — coursiv_ · 2026-11-09
- FOCS 2026 opens in NYC, jokingly called the last conference with human-proven theorems — fortnow · 2026-11-09
- NVIDIA's SIGGRAPH 2026 paper DAT delivers penetration-free contact for cloth, soft bodies and robots — zhengyiluo · 2026-11-09