Billy Qian

Biophysics PhD student @Harvard | @KempnerInst

[Billy Qian @b_qian_]

@fatihdin4en And yes, happy to discuss more offline once the conference is over! (4/)

[Billy Qian @b_qian_]

Excited to share my #NeurIPS2024 paper with @jzavatoneveth, @BenjaminSRuben, and @CPehlevan on mechanistic mismatches in data-constrained models of neural dynamics! (1/n)

[Billy Qian @b_qian_]

@fatihdin4en This isn't unique to CORNN. And overall, the point here is to show that these mismatches *can* arise in a very particular way , and under theoretically understood conditions (partial observation and non-normal teacher connectivity). (reply 3/)

[Billy Qian @b_qian_]

@fatihdin4en It's not clear why cross-validation on say, MSE of fitted activity would find the level of regularization in the plot you posted for example, where the explained variance of the fit is already effectively zero. (reply 2/)

[Billy Qian @b_qian_]

@fatihdin4en We use the same reg. strength (1e-3) across all methods. Setting a high enough regularization can of course ablate any "spuriousness", but in the trivial sense that the weights collapse towards the zero matrix, where dynamics simply decay from their initial condition.

[Billy Qian @b_qian_]

@fatihdin4en Agreed, although I should note that we motivate the single trial setup with the setting of recording from animals performing naturalistic behaviors, where repeatable trial structure isn't possible.

Thanks Fatih!

[Billy Qian @b_qian_]

@fatihdin4en In the example plots you show, the fitted activity decays monotonically to the trivial fixed point, capturing neither the oscillations nor the nonnormal transients of the teacher. The proper level of regularization one should use is a delicate balance that isn't obvious a priori.

[Billy Qian @b_qian_]

@fatihdin4en Certainly! Most of our analysis treats the ridgeless case for simplicity, but proper regularization can help ablate these effects. However, this also comes with a tradeoff in terms of actually meaningfully fitting salient features of the observed activity. (reply 1/2)

[Billy Qian @b_qian_]

Find the full paper here: https://t.co/tFFRtNoFf3 (n+1/n)

[Billy Qian @b_qian_]

If you’d like to learn more, swing by our poster on Friday (tomorrow) at 11am, East Exhibit Hall #3807, or at the NeuroAI workshop on Saturday at 3:30pm, West Ballroom B! (11/n, end)