CMC lab Computational Machinery of Cognition


The overarching research theme of our lab is understanding the computational machinery of cognitive processes. Cognition spans a wide range of functions (from perception to planning), and it is one of the most remarkable capabilities of the brain. In CMC lab we want to understand the computations underlying cognitive processes, and the biophysical machinery that implement these computations. Imagine we go to a restaurant and want to order a dish. For such decisions brain needs to take into account several factors: how much we like each option, how much we want to explore new options, how pricey they are, … . There is much computation going on in our brain to sort out each of those questions. And, at the end, all is happening in a piece of wet machinery (brain). We want to understand how our brain does those computations and how they are implemented (biophysics of computations).

We do normative and biophysical modeling (with the goal of combining them) to understand the cognitive functions; we will test these models with neural and behavioral data (in collaboration with experimental labs); and we will develop machine learning methods for multi- and cross-scale analysis of neural data to first better understand the multi-scale machinery of the brain, and second better capture the neural markers of underlying cognitive (sub-)processes and ultimately connect them to underlying computations.

At the moment we are perusing our key goal through the following research program?

  1. CMPD: Computational Machinery of Perceptual Decision-making
  2. MNBD: Multi-scale analysis of Neural and Behavioral Data
  3. CFDN: Criticality in Functional and Dysfunctional Neural systems
  4. HNNC: Hybrid Neural Networks for Cognitive Neuroscience

Computational machinery of perceptual decision-making (CMPD)

We want develop functional multi-scale models of decision processes, which is a key aspect of information processing.  We will start this development from perceptual multistability (phenomena like Necker cube and binocular rivalry), that is also a form of perceptual decision.

Read more:

We chose perceptual multistability as our starting point due to its multi-faceted richness:

  1. It is an evolutionary-preserved phenomenon.
  2. It is studied across several levels of organization (from genes to brain networks).
  3. It is a rich task to understand different aspects of neural computaitons. This includes, perceptual inference and we recently showed it is also a rich task for understanding value-based decision-making, and even cognitive control.
  4. It is broadly applicable across many species, from Drosophila to human, as well as across different sensory modalities.

Furthermore, perceptual multistability is also important from a psychiatric perspective, as it has been found to differ in a wide range of psychiatric conditions (e. g., differences in the rate of perceptual switches), So, we believe studying this phenomenon has great potential for gaining an integrative insight into a wide range of neural functions and dysfunctions.

Highlighted paper of the research program

Multistability, perceptual value, and internal foraging
Multistability, perceptual value, and internal foraging
Shervin Safavi, Peter Dayan
Neuron   ·   01 Oct 2022   ·   doi:10.1016/j.neuron.2022.07.024
Based on a large collection behavioral and neural data, we introduce a radically new perspective to the well-known phenomenon of perceptual multistability. We formulated the phenomenon as a value-based decision-making process that not only synergies with previous accounts, but also offers a more comprehensive treatment of computational and algorithmic facets of perceptual multistability.

Check here for more papers from CMC lab on this research program.

Multi-scale analysis of neural and behavioral data (MNBD)

We develop multi-modal methods for the analysis of brain data. The destination of this methodological research is to establish the connections from multi-scale neural dynamics to behavior, and ultimately information processing.

Read more:

The organization of the brain spans multiple levels (neurons, circuits, networks, etc), which are duly assessed using different measurement modalities (electrophysiology, calcium imaging, fMRI, etc). Despite extensive developments in the analysis of individual data modalities, studies that focus on joint analysis of multi-modal neural data are still scarce.

We started this research program from Neural events. Neural events are characteristics, transient, coordinated, neural activities that we can identify them in aggregated signals (e.g., local field potentials or LFPs). We want to use them as accessible neural markers of cognitive processes, that likely shed light on behaviorally relevant coordination mechanisms in the brain.

It has been shown some neural events have signature across several scales (neurons, neural populations, and large-scale networks). Moreover, they are also closely connected to behavior; for instance Sharp-Wave Ripples is one of the most studied neural events and, over two decades, it has been shown they are involved in everything from memory consolidation to offline and online planning. We want to use them as accessible neural markers of cognitive processes, that likely shed light on behaviorally relevant coordination mechanisms in the brain.

key gaps that we are trying to fill are the following:

  1. How we can detect neural events in data reliably? Thus, we will develop unsupervised machine learning methods to identify neural events.
  2. If and how neural events are coupled to behavior? Thus we will use our method(s) to investigate the occurrence of different kinds of neural events in variety of behavioral task.
  3. How rich multi-scale dynamics of neural event support the behavior? Thus, we will characterize the multi-scale signature of neural events (e.g., how different brain regions interact/communicate during each event) during variety of behavioral task.

Highlighted paper of the research program

Uncovering the organization of neural circuits with Generalized Phase Locking Analysis
Uncovering the organization of neural circuits with Generalized Phase Locking Analysis
Shervin Safavi, Theofanis I. Panagiotaropoulos, Vishal Kapoor, Juan F. Ramirez-Villegas, Nikos K. Logothetis, Michel Besserve
PLOS Computational Biology   ·   03 Apr 2023   ·   doi:10.1371/journal.pcbi.1010983
In this study we provide one of the first multi-variate methodology for investigating spike-LFP coupling, that allow us to connect neural activity to population dynamics. The method accompanied with large body of biophysical simulation, as well as analysis of prefrontal recurrent circuitry.

Check here for more papers from CMC lab on this research program.

Criticality in functional and dysfunctional neural systems (CFDN)

We will use “brain criticality hypothesis” which has the potential of bridging multi-scale neural dynamics to basic information processing capabilities, as well as behavior.

Read more:

Crudely speaking, this hypothesis states that the brain operates close to the edge of instability (e.g., a sweet spot between over-synchronization and random activity). Operating in this regime explains some key features of neural dynamics that are particularly important for a multi-scale description of the brain (e. g., scale-freeness). Furthermore, criticality has been suggested to be an optimized regime for information processing.

We extend the previous line of research in two directions.

  1. Extend it to a broader range of information processing and other biologically relevant neuronal networks.
  2. We investigate if deviation from the critical state occurs in psychiatric disorders.

Highlighted paper of the research program

Signatures of criticality in efficient coding networks
Signatures of criticality in efficient coding networks
Shervin Safavi, Matthew Chalk, Nikos Logothetis, Anna Levina
Cold Spring Harbor Laboratory   ·   14 Feb 2023   ·   doi:10.1101/2023.02.14.528465
In this study, we probe the connection between the optimality discussed in the context of criticality hypothesis of the brain (that very much tied to neural dynamics), and the optimality discussed in theories of neural computations. We examine an efficient coding network for signatures of criticality and we find signatures of criticality emerging in a network that was designed based on criticality-agnostic principles merely by optimizing the coding performance. This suggests criticality and efficient coding might be intimately related.

Check here for more papers from CMC lab on this research program.

Hybrid neural networks for cognitive neuroscience (HNNC)

Artificial neural network (ANN) have been influential for understanding the neural computations, however they suffer from biological plausibility. This is a key caveat when we want to use them to understand the computational machinery of cognitive process.

Read more:

We will exploit the capacity of ANN for implementing different kinds of computations and hybrid them as much as possible with biological neural networks. This will let us to have the goodness of both worlds.