Excited to share that the collaborative work with @mhahn29 just came out at @NatureNeuro . https://t.co/6tFgwgMHpv

We are searching for a new faculty member who works in the interface between neuroscience, AI, and computer science. This is a joint search between Departments of Neuroscience and Computer Science at UT Austin. Aplication link: https://t.co/PAPjIbuuaC

Honored to be named a Sloan Research Fellow. I am grateful for the support from my family, mentors, collaborators, mentees, and friends over the years, and for the Sloan Foundation for supporting our research. #SloanFellow https://t.co/kywFoxyxWo

Overall, the results provide a unified theory of attractive & repulsive biases reported over decades of research. OA link:

Second, biases reported in orientation perception can be explained by an efficient neural code together with a prior distribution that is nearly uniform.

Application of our theory to datasets across multiple modalities leads to several surprising observations. First, there is a major difference between the origin of the perceptual biases for circular variables and scalar variables.

We analytically derived how biases in Bayesian observers depend on prior distribution, the heterogeneity of encoding precision and loss function. This leads to a simple rule to judge the direction of biases depending on whether the prior or encoding precision changes faster.

Perceptual biases have been well-documented for many stimulus variables. It is often thought that prior expectation is the main reason underlying these biases. In this paper, we present analytical and empirical evidence against this classic view.

(20/n) According to this view, it makes sense to have multiple (rather than one) copies of head direction representation. This provides one conceptual framework to think about why there are so many brain areas in the mammalian brain that encode head direction.

(19/n) Its representation should be more modifiable compared to the cortex, which may maintain multiple hypotheses of the world. The long-term memory in the cortex can correct the thalamus if needed.

(18/n)Final remarks: my own interpretation (not in the paper): these data on the head direction cells in the thalamus is consistent with the idea that the thalamus acts as an “active blackboard” (David Mumford, 1991), which encodes the current best estimate of the world.

(17/n) The memory traces observed here may enable the system to be more robust and flexible in ambiguous environments, and when switching between environments.

(16/n) To me, the most surprising finding is the long-term memory traces of sensory cues in the HD system. Previously, I had always thought that the HD system only encodes the instantaneous estimate, with no memory of the previous states.

(15/n) Overall, these results challenge the traditional characterization of the HD system as a one-dimensional ring attractor system and suggest that the HD system contains rich structures and dynamics.

(14/n) This observation could be explained by incorporating another set of neurons into the network model to account for the difference between the “perceived” and “vestibular” angular velocity.

(13/n) Experimental observation 3: A rotating visual cue with a constant speed can cause the internal HD representation to keep rotating with that speed even after the cue is turned off. This suggests that the system recalibrates the angular velocity gain using the visual cue.

(12/n) To keep track of the HD, it is critical to accurately integrate the angular velocity. While this process is thought to be vestibular-driven, here we find that visual information can also play an important role (see below):

(11/n) Incorporating a Hebbian-like plasticity mechanism in a continuous attractor model to modify the weights between the ADN HD cells and an additional population of neurons encoding HD based on visual cues can explain these experimental observations.

(10/n) Past visual experience can cause systematic shifts of the HD representation in darkness, presumably through these memory traces.

9/n Experimental observation 2: the ADN HD cells store memory traces of past visual experiences for at least minutes. These memory traces appear to be stored in the form of HD-specific gain— directions that are aligned with the past visual cues are encoded with a larger gain.