Kanaka Rajan
@kanakarajanphd.bsky.social did:plc:nemx43ymm6toi4fqil4gqbev
Associate Professor at Harvard & Kempner Institute. Applying computational frameworks & machine learning to decode multi-scale neural processes. Marathoner. Rescue dog mom. https://www.rajanlab.com/
Bei Bluesky seit
- Follower
- 3.785
- Folgt
- 248
- Posts
- 61
- Tage in Ranglisten
- 0
Zusammenfassung
Kanaka Rajan (@kanakarajanphd.bsky.social) ist seit 11 September 2023 auf Bluesky. Das Konto hat 3.785 Follower, folgt 248 Konten und hat 61 Posts veröffentlicht.
Ranglisten-Platzierungen
Die besten Tage, an denen Posts dieses Kontos in Debyas Tagesliste kamen.
Dieses Konto war noch in keiner Tagesrangliste.
Zuletzt gelikte Posts
Öffentliche Likes aus dem Bluesky-Netzwerk.
📄 Paper: arxiv.org/abs/2609.38081 🌐 Page: ann-huang-0.github.io/Hessian-null... 💻 Code: github.com/Ann-Huang-0/hessian-null-space-continuation w/ my awesome team 🌟 @neurostrow.bsky.social, @hansonlu.bsky.social @wtredman.bsky.social, @leokoz8.bsky.social, @kanakarajanphd.bsky.social (12/12)
Post anzeigenOverall, HNC finds diverse mechanisms with matched function around a single trained network. This opens the door to mechanistic analysis of solution spaces, to separating task demands from optimizer bias, and to a principled basis for reward auditing, model merging, editing and fine-tuning (11/12)
Post anzeigen📐 HNC also measures loss-landscape geometry. The Hessian captures the local geometry around one network; the null space walk tracks how it changes along the way. Bigger models have more flat directions, while harder tasks have fewer. The directions HNC follows get stiffer on harder tasks. (10/12)
Post anzeigenHNC can also uncover reward hacking! In the AI Safety Gridworld boat race, we start with policies that correctly follow the track. Preserving proxy reward (+3 for entering an arrow tile clockwise), HNC walks find the known exploit: hopping on and off one tile without net clockwise progress (9/12)
Post anzeigenApplied to RL, HNC searches for different policies while preserving a reward surrogate. An odor-tracking agent shifts from zigzagging across a plume to smoothly following its edge—a distinct navigation strategy at comparable return 🪰(8/12)
Post anzeigen
Blockierte Konten
Blockierungen sind auf Bluesky öffentlich.
Dieses Konto blockiert niemanden.