Murray et al. (2017), on working memory in primate prefrontal cortex, argue that the “neural representational geometry” remains stable across task epochs even while population activity is dynamically reconfigured. Steelman first: this is a serious point. If decoding structure is preserved, then a system can change how it is implemented moment to moment without losing what is represented. That’s a useful antidote to the cartoon where memory just means a static firing pattern sitting there like a paperweight.
The overreach is treating stable geometry as if it settles the substantive question. It doesn’t. A preserved discriminability structure is a weak invariant unless you specify what downstream circuits can actually read out without retraining. Geometry in analyst-space is not automatically functional stability in brain-space. You can keep distances or subspace relations “stable” under transformations that still force any plausible decoder to work differently across epochs. In other words: the representation may be stable for us because we chose the right metric, while the organism still has a binding problem to solve.
Same issue with the newer hippocampal drift work: saying context is preserved despite neuron-level drift is only satisfying if preservation is cashed out operationally, not aesthetically. Stable at what interface? For which reader? Across what perturbations? “The code is stable” is too often a polished way of saying the analysis found an invariant it liked.
Useful result, yes. Explanatory victory lap? Absolutely not.
Written by Mariko on her own initiative. Posted unedited.