Jenelle Feather

1/6 I usually don’t comment on these things, but @RylanSchaeffer et al.'s paper contains enough misconceptions that I thought it might be useful to address them. In short, effective dimensionality is not the whole story for model-brain linear regression, for several reasons: 🧵👇 (1) The spectral theory due to @canatar_a @jenellefeather @s_y_chung is misrepresented in this paper, and crucially relies on model alignment with brain data (this is already evident from their equations). In fact, they conclude that effective dimensionality alone doesn’t fully predict neural data. (2) Even in this paper’s own Figure 2 (top left panel), networks with low participation ratios, like SRNN, still achieve high neural predictivity, already indicating that participation ratio is not a unilateral predictor of whether a model will match the brain via linear regression. (3) Beyond MEC, this lack of a trend with effective dim. is also the case when looking at models in their match to macaque IT or human OTC. For example, Colin Conwell @_jacobprince_ @talia_konkle et al.’s very thorough work: biorxiv.org/content/10.1101/… shows that effective dim isn’t related to linear prediction of human OTC (cf. their Figure 5). Moreover, in macaque IT, as @apurvaratan & others have seen, if you include very predictive models of IT responses, the effective dim to linear predictivity trend also doesn’t hold. (@jenellefeather & others don’t see this trend holding up in matches to auditory cortex either.) (4) Furthermore, the models they cite participation ratios for, were also compared against several non-fitted metrics (like RSA, score distributions, simpler-than-linear mappings, etc.), which found similar conclusions that matched linear regression results, across MEC, IT, and auditory cortex.