Excited to share my work with @EngelTatiana, out now in @NatNeuro! We show that RNNs use low-dimensional latent circuit mechanisms for cognitive tasks. We find that context-dependent decisions in both RNNs and PFC arise from latent inhibitory mechanisms.

(5/5) These results demonstrate that modeling the low-dimensional recurrent mechanisms underlying task execution is critical for identifying low-dimensional behaviorally relevant representations.

(4/5) A latent circuit model of PFC responses similarly reveals low-dimensional projections of neural activity in which irrelevant stimuli are suppressed. Larger representations of irrelevant stimuli on error trials confirms the behavioral relevance of the latent circuit.

(3/5) The orthonormal embedding allows validating the latent circuit mechanism via low-dimensional projections of activity and low-rank perturbations of connectivity. In contrast to previous studies, we find representations of stimuli are suppressed when they are irrelevant.

(2/5) To discover these latent mechanisms from high-dimensional responses, we developed the latent circuit model which simultaneously learns a low-dimensional recurrent neural network and an orthonormal embedding to jointly explain heterogeneous neural responses.

We are excited to present a Brain-Wide Map of neural activity during a complex decision-making behavior, a product of the monumental effort of 22 labs: 547 Neuropixels recordings in 267 brain regions from 115 mice collected in 11 labs 🐭 (1/8) https://t.co/5W0RqM6xpB

(4/4) How should we understand this relationship in neural populations with mixed selectivity to multiple task variables? Theoretical work in RNNs demonstrates latent circuit mechanisms may drive low-dimensional dynamics in heterogeneous networks through low-rank connectivity.

(3/4) In mammalian entorhinal cortex, manifold analyses reveal toroidal structure stable across environments and behavioral states. This rigidity is consistent with continuous attractor models in which circuit connectivity is spatially arranged on a two-dimensional manifold.

(2/4) The correspondence between the neural manifold and underlying circuit structure is conspicuous in cases in which the geometry of neural responses mirrors their spatial layout in the brain, such as in the fly navigational system.

(1/4) Excited to share our perspective article with @EngelTatiana and @MGENK out today in Nature Reviews Neuroscience! We highlight work which unifies the neural manifold and circuit perspectives on cognition and argue that they are inseparable.

Thrilled to share our paper "Intrinsic timescales in the visual cortex change with selective attention and reflect spatial connectivity" now out @NatureComms! https://t.co/HLWLDYcp0h We added interesting new analyses to our preprint version (thanks to the reviewers!) #tweeprint👇

Excited to share our recent work on how selective attention modulates intrinsic timescales of local spiking activity in visual cortex and its underlying network mechanisms https://t.co/EUNcsMJlki #tweeprint + some background on intrinsic timescales👇

Excited to share our new preprint on how network dynamics and structured connectivity jointly define the spatial and temporal profiles of neural correlations. Work w/ @roxana_zeraati, @SelfOrgAnna, @EngelTatiana: https://t.co/Vy4YQhOqHf Here is a #tweeprint:

In decision-making tasks excitatory and inhibitory neurons show selectively for an upcoming choice. @EngelTatiana, @anne_churchland, and I explore the effects of inhibitory choice selectivity in decision making circuits using mean-field and RNN models.

Seemingly in conflict with the inhibitory mechanism, a decoder finds projections of RNN activity in which irrelevant stimuli are not suppressed. We show that these representations do not causally drive behavior. Thus, decoding results cannot be interpreted as causal mechanisms.

The latent circuit is a mechanistic model that predicts interpretable behavioral changes under connectivity perturbations. The embedding Q allows us to translate these perturbations onto RNN and causally test the circuit mechanism via patterned connectivity perturbations.

Latent circuit inference identifies low-dimensional latent connectivity structure which supports the behavioral task performance. This latent connectivity is revealed in RNN by a change of basis given by the embedding Q.

We applied latent circuit inference to RNNs trained to perform context-dependent decision-making and found a latent inhibitory mechanism for flexibly switching between alternative stimulus-response mappings.

The latent circuit model captures mechanistic interactions between task variables and their heterogeneous mixing in single neurons via the embedding Q. We simultaneously infer the latent circuit connectivity and the embedding matrix from neural responses.

Excited to share our bioRxiv preprint with @EngelTatiana. Cortical neurons show mixed selectivity for multiple task variables. We develop a latent circuit model for inferring interpretable circuit mechanisms from heterogeneous neural responses. https://t.co/KivuC5bOnq