Arnav Raha
@pantamallion.bsky.social
Neuroscience PhD student at Columbia University | UC Berkeley alum | Studying RL mechanisms for vocal learning (+ fan of animals 🐋🦇🦜🦋🦑 🦎🦕)
🚂 see y’all in Montreal 😊 #cosyne2025
March 26, 2025 at 2:13 PM
🚂 see y’all in Montreal 😊 #cosyne2025
The final major question: is dopamine signaling pure error (just mismatch from the goal) or an actual prediction error (RL-style)? By regressing dopamine on past renditions, we see it integrates ~14 past attempts, not just the current one—just like a prediction error! (6/7)
March 17, 2025 at 5:48 PM
The final major question: is dopamine signaling pure error (just mismatch from the goal) or an actual prediction error (RL-style)? By regressing dopamine on past renditions, we see it integrates ~14 past attempts, not just the current one—just like a prediction error! (6/7)
But does dopamine actually shape learning? We used canonical correlation analysis to find the direction in song space that dopamine "rewards." Sure enough, future song shifts in that direction! Even more intriguingly—dopamine predicts future changes better than past ones. (4/7)
March 17, 2025 at 5:48 PM
But does dopamine actually shape learning? We used canonical correlation analysis to find the direction in song space that dopamine "rewards." Sure enough, future song shifts in that direction! Even more intriguingly—dopamine predicts future changes better than past ones. (4/7)
The results? Dopamine tracks relative performance! When a bird sings better than recent attempts, DA spikes. When it sings worse, DA drops. This relationship scales neatly with magnitude—see my favorite figure below! This aligns perfectly with an RL reward prediction error framework. (3/7)
March 17, 2025 at 5:48 PM
The results? Dopamine tracks relative performance! When a bird sings better than recent attempts, DA spikes. When it sings worse, DA drops. This relationship scales neatly with magnitude—see my favorite figure below! This aligns perfectly with an RL reward prediction error framework. (3/7)
Came across another 2007(?) vs 2020! Though I think I can do a bit better now after having worked with them 😜
January 2, 2025 at 4:51 AM
Came across another 2007(?) vs 2020! Though I think I can do a bit better now after having worked with them 😜
