Churchland lab

It’s been a while since we’ve updated the ORACLE! We are still reading papers on the preprint server, but our ability to write about them got a bit slowed down: COVID happened, and then we moved to UCLA! We’ll try to ramp back up now that we are getting settled. The paper below is one that I read with James Roach, a postdoc in my lab, and our conversations about it shaped my thinking.

The paper: Attractor dynamics gate cortical information flow during decision-making; Arseny Finkelstein, Lorenzo Fontalan, Michael Economo, Nuo Li, Sandro Romani & Karel Svoboda.

This crew aimed to uncover how sensory information can have a differential impact on actions depending on when in the course of a decision it is presented. They developed this approach to tackle this problem: they presented trained mice with distractors (photostimulation to primary somatosensory cortex) at different times during a delay period that preceded a response period. Behaviorally, stimuli presented earlier had the biggest impact, indicating that later stimuli were “gated out” or, were somehow prevented from impacting the neural activity that drove the licking response. To determine the mechanism of this gating, the authors measured neural activity in a region of the frontal cortex, the anterior lateral motor cortex (ALM). These measurements demonstrated that neural activity likewise became more robust to distractors over time. To me, this was the main take-home: the sensory stimuli still impacted ALM activity, but the choice-related dynamics became progressively more robust over the trial so that these same stimulus-driven deflections had a smaller effect.

In an artificial network trained up to generate outputs matching the mean responses of neurons he recorded in ALM. In the artificial network, attractor dynamics stabilized the activity of the units, making them more robust to perturbations. A key feature of this network was a “ramping input”, based on some observations in the data.

I liked 3 things about this paper. First, the behavioral paradigm is good: the diminishing effect of distractors over time is very robust. Also, this diminishing behavioral effect stands in contrast to the non-attenuation of the distractors in either S1 or ALM. Second, the data analysis was compelling. Specifically, the authors used a targeted dimensionality reduction approach to interpret the activity of area ALM in light of both sensory and movement planning signals. This analysis made it possible to compare ALM and S1 despite differing sensory responses in those two areas. Finally, the inclusion of an artificial network was an important addition: the network was trained to match responses of ALM neurons and naturally exhibited many of the same properties as the actual neurons. Reverse-engineering of this network made it possible to propose a mechanism: specifically, that the network became more robust to perturbations over time because a ramping increase in firing rate separated distinct attractor basins more clearly. Taken together, these highlight a potential mechanism for an interesting cognitive phenomenon and connect in an interesting way to previous literature on this subject, mainly in non-human primates.