The estimation of movement in the visual system is one of the best studied examples of computation in the brain, with results dating back to the 1800s. There is work both on neurons and behavior, and on organisms ranging from insects to primates, with surprisingly common themes. One idea is that the brain’s algorithms for motion estimation are optimized, allowing for estimates that are as accurate or reliable as possible. The problem is that the predictions depend on the underlying joint statistics of movies and motion in the world, and on the relevant noise levels. In the absence of independent measurements of these quantitates, a theory of optimal estimation just becomes a model with parameters to be fit to the data. My experimental colleagues have developed a specialize instrument that provides a “fly’s eye view” of the visual world, allowing us to sample the joint distribution of movies and motion in natural environments. Theory tells us how to construct the form of optimal motion estimators, quantitatively, as averages over these estimates. We also know the irreducible noise levels in the fly’s retina, we can record the responses of motion sensitive neurons deep in the fly’s brain, and we can map related these computations to the connectome. This defines a program of comparing optimal estimates with the real computations done by the brain, in detail and with no free parameters. We are far form done but I will described our progress.

Pizza will be served for lunch!

“Entropy Rate of Meaningful Discourse”

Misha Tsodyks, Ph.D.

Misha will discuss the attempts to leverage Large Language Models to evaluate information content of meaningful discourse by separating meaning from phrasing.

Pizza will be offered for lunch!

Katharina Duecker, Ph.D.

Title: The next-generation Human Neocortical Neurosolver (HNN): biorealistic cell models reveal the importance of dendritic spiking dynamics in interpreting human EEG/MEG signals

“Flexible Error Monitoring in Context-dependent Behavior” 

Atsushi Kikumoto, Ph.D.

Research Associate in Cognitive and Psychological Sciences, Brown University

Join us for the first talk of the CCBS Flash Talk series! These talks will be interactive with open discussion, and small-group conversation.

Sheridan Feucht, Ph.D. Candidate, Northeastern University

What is a word to an LLM? Individual tokens are often semantically unrelated to the meanings of the words/concepts they comprise, meaning that LLMs must build up representations of word meaning in-context that are separate from token identities. In this talk, I show how these multi-token words can be used as a wedge to separate conceptual representations from literal token representations. I describe our “dual-route theory of induction”, which argues that models copy literal token information in parallel with “fuzzy” word representations, and show that this “fuzzy” concept induction is vital for other tasks, like translation. Finally, I describe our recent work analyzing the geometry of these independent subspaces, which suggests that they are likely useful for more than just copying.

This talk will highlight recent advances in applying deep learning and digital twin models to neuroscience, and how these approaches inform artificial intelligence. By examining the responses of V1 and V4 neurons to a large dataset of natural images (~40,000), we find that their neural codes display remarkable complexity and diversity, challenging long-standing assumptions. Using deep learning–based digital twins of these neurons, we conduct both classical neurophysiological experiments in silico and novel computational studies to probe their selectivity for natural image features, the neural mechanisms that support such selectivity, and the computational constraints that shape their topological organization. These insights provide a deeper understanding of how concepts are learned and composed in the nervous system, with potential implications for building more flexible and interpretable AI systems.

Please join ANCOR (AI, Neuro, CogSci Research talks) on Monday 3/17 at 11am featuring Sam Lippl (PhD student at Columbia University/Zuckerman Institute). 

Zoom: https://brown.zoom.us/j/93766900152

The success of Classical (i.e., symbolic) linguistic theories suggests that at least in the linguistic domain, the mind constructs structured symbolic representations and processes them with rules of symbol manipulation. However, modern Connectionist (i.e., neural) models may provide an alternative foundation for linguistic computation: Connectionist models often lack built-in mechanisms for Classical representation and processing, yet they perform impressively on a wide array of linguistic tasks. Connectionist models may not challenge the Classical approach if they implement Classical computers. This is no mere theoretical possibility: when trained on simple symbol manipulation tasks, small Connectionist models develop structured symbolic representations, a key aspect of Classical computation. This raises the possibility that Connectionist models trained on natural data also develop such representations. However, structured representation is not sufficient for Classical computation. The processing of the structured representations by the Connectionist models must involve abstract symbol manipulation of the sort that Classical theories posit. Else, Connectionist models may still challenge the Classical account of human linguistic capacities. We study this question by testing if Connectionist models trained on simple symbol processing tasks develop Classical processing mechanisms. We find evidence suggesting that Connectionist models that succeed on such tasks do implement Classical models. To the extent that such findings generalize to models trained on naturalistic data, such a result would suggest that modern Connectionist models do not challenge Classical theories of human language.

Hi friends,

Next Monday we are hosting another ANCOR (AI, Neuro, Cog Research talks) speaker! Ekdeep Lubana (Postdoc, Harvard) joins ANCOR to present his work Dynamics of Concept Learning and Emergent Abilities in Neural Networks.

Date: Monday, Feb 10, 11am
Location: CIT 477, 115 Waterman St.

All are welcome!
Mikey + Aalok

https://brown-ancor.github.io/

Ciana Deveau (Brown/NIH) joins ANCOR (AI, Neuro, Cog Research talks) to present”Recurrent cortical networks encode natural sensory statistics via sequence filtering”.

Speaker: Eghbal Hosseini (Postdoc, MIT)

Title: Large language models implicitly learn to straighten neural sentence trajectories to construct a predictive representation of natural language

A predictive coding perspective on oscillatory traveling waves

Andrea Alamia, Centre de Recherche Cerveau et Cognition (CerCo), CNRS, Université de Toulouse, Toulouse, France

This talk presents a few studies that aim to interpret oscillatory travelling waves in the predictive coding framework. In the first part, I’ll introduce a simple model of the visual cortex based on predictive coding mechanisms, in which physiological communication delays between levels generate alpha-band rhythms. Interestingly, these oscillations propagate as traveling waves across levels, both forward (during visual stimulation) and backward (during rest). Remarkably, experimental EEG data matched the predictions of our model. In the second part of the talk, I’ll present two studies that indirectly investigate the link between predictive coding mechanisms and traveling waves experimentally: the first one investigates the effect of a powerful psychedelics drug, N,N, dimethyltryptamine (DMT), on alpha-band oscillations, and the second one interprets the pattern of oscillatory traveling waves in schizophrenic patients in the light of Predictive Coding. In the last part of the talk, I will show some (very) preliminary results on a statistical learning paradigm that directly explores the link between traveling waves and predictive coding processes.

Large Language Models vs Human Brain: mapping and decoding the language code in neural systems

Jean-Rémi King, Ph.D.

ShiNung Ching, Associate Professor Electrical & Systems Engineering,
Washington University in St. Louis

A fundamental challenge in computational neuroscience has been the translation of multimodal data into formal mathematical and computational models that can reveal biophysical mechanisms in neural circuits and their connection to behavior. In this presentation, I will describe our recent efforts in this domain, focusing on extracting from data the generative dynamics that give rise to overt observations of brain activity. Specifically, I will describe how we have adapted tools from Bayesian filtering and algorithmic optimization toward the problem of parametrically learning high-dimensional, biophysically interpretable models of network interactions involving hundreds to thousands of neural populations. These techniques place a particular emphasis on model-building at the level of individuals, which in turn provides leverage on revealing idiosyncrasies in brain mechanisms. In this regard, I will highlight two ways in which we are leveraging the obtained models. First, I will discuss our newly developed methods to directly interrogate the intrinsic dynamics within models, toward assessing topological similarity across individual (brain) dynamics and the functional salience thereof. Second, I will describe how we are using models to predict input-output relationships within brain networks and their responses to exogenous, causal perturbations. In addition to basic mechanistic insights, these approaches enable us to design brain stimulation protocols that are tailored to individuals and defined in terms of dynamical targets that can be linked to specific functional endpoints; to conclude I will briefly describe ongoing work to validate this latter premise.

Deconstructing human reinforcement learning

Anne Collins, Ph.D. Associate Professor of Psychology at the University of California, Berkeley

Reinforcement learning frameworks have contributed tremendously to our better understanding of learning processes in brain and behavior. However, this remarkable success obscures the reality of multiple underlying processes that support humans’ unique flexibility and adaptability. In this talk, I will show that not accounting for such underlying processes in computational cognitive modeling weakens the generalizability and interpretability of findings, with important consequences in neuroscience, developmental, clinical research. I will present multiple approaches to disentangle the multiple processes that support flexible learning, including episodic and working memory processes. This works highlights the importance of studying learning as a multi-dimensional phenomenon that relies on multiple separable but inter-dependent computational mechanisms. Insights from how the brain implements learning is essential to informing generalizable, interpretable cognitive modeling. 

Building Performant and Brain-Like Recurrent Models from Neurons and Astrocytes

Leo Kozachkov, Ph.D., Postdoctoral Associate, McGovern Institute for Brain Research, MIT

The brain’s ability to perform challenging tasks is facilitated by its many inductive biases—hardwired biological features that predispose it to process information in certain ways over others. These features include anatomically distinct brain areas, as well as specialized cell types such as neurons and glia. Inductive biases grant the brain computational powers that currently surpass artificial intelligence systems in many domains. In this seminar, I will cover two recent avenues of research that leverage the brain’s inductive biases to build highly performant, recurrent artificial networks.

In the first half of my talk, I will discuss recent progress in understanding the computational role of different cell types. I will focus on neuron-glial interactions. An intriguing fact is that most human brain cells are not neurons, but rather glia. There is mounting experimental evidence suggesting that astrocytes, a specialized type of glial cell, play a significant role in learning, memory, and behavior. However, our theoretical understanding is lagging far behind. I will cover recent work that aims to bridge this gap by relating dynamical, energy-based neuron-astrocyte networks to powerful AI models such as Modern Hopfield Networks and transformers.

In the second half of my talk, I will discuss how and why the brain maintains a balance between flexibility and stability through “dynamic attractors”, which are reproducible patterns of neural activity in response to (potentially time-varying) stimuli. This work reveals an unexpected and useful theoretical link between dynamic attractors and modularity. Specifically, recurrent neural networks with dynamic attractors can be combined into large, modular “networks of networks”, reminiscent of the brain’s macroscopic organization, in ways that provably preserve stability. These higher-order, stable networks can then be optimized for state-of-the-art performance on benchmark sequential processing tasks, demonstrating that dynamic stability is a useful inductive bias for building brain-like performant recurrent models.

“Bridging scales in intelligent systems– from octopus skin to mouse brain”

Leenoy Meshulam Ph.D.

Swartz Theory Fellow, University of Washington

For an animal to perform any function, millions of neurons in its nervous system furiously interact with each other. Be it a simple computation or a complex behavior, all biological functions involve many individual units. A theory of function must specify how to bridge different levels of description at different scales. For example, to predict the weather, it is irrelevant to follow the velocities of every molecule of air. Instead, we use coarser quantities of aggregated motion of many molecules, e.g., pressure fields. Statistical physics provides us with a theoretical framework to specify principled methods to systematically ‘move’ between descriptions of microscale quantities (air molecules) to macroscale ones (pressure fields). Can we hypothesize equivalent frameworks in the nervous system? How can we use descriptions at the level of neurons and synapses to make precise predictions of activity and behavior? My research group will develop theory, modeling, and machine learning tools to discover generalizable forms of scale bridging across species and behavioral functions. In this talk, I will present lines of previous, ongoing, and proposed research that highlight the potential of this vision. I shall focus on two seemingly very different systems: mouse brain neural activity patterns, and octopus skin cells activity patterns. In the mouse, we reveal striking scaling behavior and hallmarks of a renormalization group- like fixed point governing the system. In the octopus, camouflage skin pattern activity is reliably confined to a (quasi-) defined dynamical space. Finally, I will touch upon the benefits of comparing across animals to extract principles of multiscale function in the nervous system, and propose future directions to investigate how macroscale properties, such as memory or camouflage, emerge from microscale level activity of individual cells. 

“Successes and failures of machine learning models of sensory systems”

Jenelle Feather, Ph.D. 

Flatiron Research Fellow at the Center for Computational Neuroscience

The environment is full of rich sensory information. Our brain can parse this input, understand a scene, and learn from the resulting representations. The past decade has given rise to computational models that transform sensory inputs into representations useful for complex behaviors such as speech recognition or image classification. These models can improve our understanding of biological sensory systems and may provide a test bed for technology that aids sensory impairments, provided that model representations resemble those in the brain. In this talk, I will discuss my research program, which aims to develop methods to compare model representations with those of biological systems and to use insights from these methods to better understand perception and cognition. I will cover experiments in both the auditory and visual domains that bridge between neuroscience, cognitive science, and machine learning. By investigating the similarities and differences between computational model representations and those present in biological systems, we can use these insights to improve current computational models and better explain how our brain utilizes robust representations for perception and cognition. 

“The Relational Bottleneck and the Emergence of Cognitive Abstractions”

Taylor Webb, Ph.D.

Human cognition is characterized by a remarkable ability to transcend the specifics of limited experience to entertain highly general, abstract ideas. Efforts to explain this capacity have long fueled debates between proponents of symbol systems and statistical approaches. In this talk, I will present an approach that suggests a novel reconciliation to this long-standing debate, by exploiting an inductive bias that I term the relational bottleneck. This approach imbues neural networks with key properties of traditional symbol systems, thereby enabling the data-efficient acquisition of cognitive abstractions, without the need for pre-specified symbolic representations. I will also discuss studies of perceptual decision confidence that illustrate the need to ground cognitive theories in the statistics of real-world data, and present evidence for the presence of emergent reasoning capabilities in large-scale deep neural networks (albeit requiring far more training data than is developmentally plausible). Finally, I will discuss the relationship of the relational bottleneck to other inductive biases, such as object-centric visual processing, and consider the potential mechanisms through which this approach may be implemented in the human brain.

Gabriel Kreiman, Ph.D.

Professor, Harvard University, Children’s Hospital Boston

What information do neurons along the ventral visual cortex represent? Exhaustively examining all possible images is empirically impossible. Therefore, to investigate stimulus preferences, investigators have used a combination of intuitions derived from previous studies, natural stimulus statistics, and serendipitous findings. Here I will describe an approach to uncover what neurons want using a real-time, unbiased, systematic algorithm based on computational models of the ventral visual cortex. We use a generative deep neural network as a vast and diverse hypothesis space. A genetic algorithm searches this space for stimuli guided by neuron preferences. We show that this approach can rapidly generate synthetic images that trigger high activations, both in model units as well as in real neurons, in many cases even higher activations than those elicited by large numbers of hand-picked natural stimuli or images derived from conventional approaches. This approach forces us to revisit how we think about neural coding in the ventral visual cortex. I will also show the results of psychophysics experiments where humans are asked to describe the images that trigger high activation patterns in inferior temporal cortex neurons, reinforcing the notion that neurons in the ventral visual cortex represent complex visual features but not semantic categories. Finally, I will show that similar conclusions can be drawn by scrutinizing the representations in artificial neural networks as coarse approximations to the processing steps along the ventral stream.

Please join us for a seminar with Larry F. Abbott, Ph.D.

William Bloor Professor of Theoretical Neuroscience and Professor of Physiology and Cellular Biophysics (in Biological Sciences)

Principal Investigator at Columbia’s Zuckerman Institute

Co-director of Columbia’s Kavli Institute for Brain Science

Title: “Modeling the Navigational Circuitry of the Fly”

Abstract: Navigation requires orienting oneself relative to landmarks in the environment, evaluating relevant sensory data, remembering goals, and convert all this information into motor commands that direct locomotion. I will present models, highly constrained by connectomic, physiological and behavioral data, for how these functions are accomplished in the fly brain.

The Carney Center for Computational Brain Science and the Brainstorm Program is organizing a two-week computational modeling workshop with a focus on computational modeling of cognition, behavior, and brain/behavior relationships. Workshop attendees will learn the basic tools for understanding, developing, and applying models to brain science questions, and have the opportunity to apply these techniques in a novel dataset.

Week 1 will consist of workshops and live tutorials, including daily lectures spanning basic to advanced topics, accompanied by hands-on coding tutorials. Attendees will learn the basic tools for understanding, developing and applying computational models, with a focus on hypothesis testing, quantitative fitting, bayesian methods, and model checks and comparisons. Additionally, advanced modeling sessions will provide a deeper theoretical understanding and application of complex modeling techniques.

During Week 2, participants will have the opportunity to work in teams to apply these skills to analyze a real dataset provided by the organizers, with potential for novel discoveries. Prizes will be awarded for models with the most predictive power, rigor, creativity, and innovation.

For details on last years’ workshops and modeling competition, visit the Center for Computational Brain Science website. Previous syllabi are available here. We will cover most of the same basic topics, with a few tweaks and additions (based on participant input and guest speakers).

Intended Audience: This workshop is open to the members of the Brown community, and is designed for researchers across fields, backgrounds and levels of experience: computation “novices” with no experience and those with more computational experience who may want to augment their toolkit with advanced approaches to parameter estimation or specific classes of models. Although there is no computational experience required, those with modeling backgrounds will still benefit from the advanced modules, and will have the opportunity to learn new skills and state-of-the-art computational approaches.

Maximum number of participants: Participation is limited to 20, but we do keep a waitlist.

Register here.

Organizers: Andra Geana, Debbie Yee, Alana Jaskir, Michael Frank

Please mark your calendars for the next MRF/BNC Users meeting, which will be Monday, March 13, at noon in the Carney Innovation Zone at 164 Angell St. We hope to see you in person but we will again have a Zoom option for those wishing to tune in remotely.

This month, Haley Keglovits and Apoorva Bhandari will co-present work from their on-going project: “Task structure shapes the geometry of control representations in PFC”.

This meeting will be streamed over Zoom for those unable to participate in-person.
Lunch will be provided.

Please RSVP for this event to help us gauge attendance and cater to any food restrictions you may have.

RSVP: https://forms.gle/XEM9PMmzCaKs2f6eA

Theory of the Multiregional Neocortex: Large-scale Neural Dynamics and Distributed Cognition

Center for Computational Brain Science Seminar Series: “Sequences and modularity of dynamic attractors in inhibition-dominated neural networks”

Carina Curto, Ph.D.
Professor
Department of Mathematics
Pennsylvania State University

Abstract: Threshold-linear networks (TLNs) display a wide variety of nonlinear dynamics including multistability, limit cycles, quasiperiodic attractors, and chaos. Over the past few years, we have developed a detailed mathematical theory relating stable and unstable fixed points of TLNs to graph-theoretic properties of the underlying network. These results enable us to design networks that count stimulus pulses, track position, and encode multiple locomotive gaits in a single central pattern generator circuit.

Title:  Reverse engineering neural control of behavior in Hydra

Join Carney’s Center for Computational Brain Science (CCBS) on June 10 for a seminar featuring Adam Calhoun, Ph.D., postdoctoral research associate at Princeton University.

Abstract

Animals produce behavior by responding to a mixture of cues that arise both externally (sensory) and internally (neural dynamics and states). These cues are continuously produced and can be combined in different ways depending on the needs of the animal. However, the integration of these external and internal cues remains difficult to understand in natural behaviors. To address this gap, Calhoun has developed an unsupervised method to identify internal states from behavioral data, and he has applied it to the study of a dynamic social interaction. During courtship, Drosophila melanogaster males pattern their songs using cues from their partner. This sensory-driven behavior dynamically modulates courtship directed at their partner. Calhoun uses his unsupervised method to identify how the animal integrates sensory information into distinct underlying states. He then uses this to identify the role of courtship neurons in either integrating incoming information or directing the production of the song, roles that were previously hidden. Additionally, Calhoun shows how song is produced by a diverse range of visual cues whose importance changes depending on behavioral context, and he identifies the visual neurons that send this information from the eye into the brain. Calhoun’s results reveal how animals compose behavior from previously unidentified internal states, a necessary step for quantitative descriptions of animal behavior that link environmental cues, internal needs, neuronal activity and motor outputs.

Michael S. Goodman ’74 Memorial Seminar Series

Speaker: Sebastian Musslick, Ph.D. student, Princeton University

Title: On the Rational Bounds of Human Cognition

Abstract: Humans are remarkably limited in the number of tasks they can execute simultaneously. This limitation is not only apparent in daily life, it is also a universal assumption of most theories of human cognition. Yet, a rationale for why the human brain is subject to this constraint remains elusive. In this talk, I will draw on insights from neuroscience, psychology and machine learning to suggest that limitations in the brain’s ability to multitask result from a fundamental computational dilemma in neural architectures. Through graph-theoretic analysis, neural network simulation and behavioral experimentation, I will demonstrate that neural systems face a tradeoff between learning efficiency (promoted through the shared use of neural representations across tasks) and multitasking capability (achieved through the separation of neural representations between tasks). Theoretical analyses show that it can be optimal for a neural system to prioritize efficient learning of single tasks at the expense of its ability to execute them simultaneously, across a broad range of conditions. These results suggest that our inability to multitask reflects a rational solution to a fundamental computational dilemma faced by neural architectures. I will demonstrate that this tradeoff can explain a variety of behavioral and neural phenomena related to human multitasking and conclude by outlining consequential computational dilemmas that may help explain other, seemingly irrational constraints on human cognition.

Please join the Carney Institute for Brain Science for a special seminar featuring Kanaka Rajan, Ph.D., assistant professor at Icahn School of Medicine at Mount Sinai.

Brown authentication is required.

Join Carney’s Center for Computational Brain Science (CCBS) on May 25 for a seminar on “Extracting structure from high-dimensional neural data,” featuring Carsen Stringer, computational neuroscientist and group leader at the Howard Hughes Medical Institute Janelia Research Campus.

Stringer completed her postdoctoral work with Marius Pachitariu and Karel Svoboda at Janelia, and her Ph.D. work with Kenneth D. Harris and Matteo Carandini at University College London. She develops tools for understanding high-dimensional visual computations and neural representations of behavior.

Abstract

Large-scale neural recordings contain high-dimensional structure that cannot be easily captured by existing data visualization methods. We therefore developed an embedding algorithm called Rastermap, which captures highly nonlinear relationships between neurons, and provides useful visualizations by assigning each neuron to a location in the embedding space. Compared to standard algorithms such as t-SNE and UMAP, Rastermap finds finer and higher dimensional patterns of neural variability, as measured by quantitative benchmarks. We applied Rastermap to a variety of datasets, including spontaneous neural activity, neural activity during a virtual reality task, widefield neural imaging data during a 2AFC task, artificial neural activity from a bipedal robot simulation, and neural responses to visual textures. We additionally found that texture identity could be decoded from these neural responses, but that the neural representations of visual texture differed from artificial neural network representations.

The Interdisciplinary Training in Computational, Cognitive, and Systems Neuroscience (ICoN) is a pre-doctoral program in computational cognitive neuroscience. Funds from this program will support the training of advanced pre-doctoral candidates who are capable of applying a combination of empirical and theoretical approaches that decisively addresses their scientific questions about the mind and brain. 

On May 7 at 3:30 p.m., join the PIs and current students to learn about the ICoN training program and how to apply to join the next cohort of students.  

Applications for this year’s program are due May 21, 2021. 

Please join Carney’s Center for Computational Brain Science (CCBS) on March 8 for a special seminar on “Untangling brain-wide current flow using neural network models,” featuring Kanaka Rajan, assistant professor of neuroscience at the Icahn School of Medicine at Mount Sinai and the Friedman Brain Institute.

Abstract:

The Rajan Lab designs neural network models constrained by experimental data, and reverse engineers them to figure out how brain circuits function in health and disease. Recently, we have been developing a powerful new theory-based framework for “in-vivo tract tracing” from multi-regional neural activity collected experimentally.

We call this framework CURrent-Based Decomposition (CURBD). CURBD employs recurrent neural networks (RNNs) directly constrained, from the outset, by time series measurements acquired experimentally, such as Ca2+ imaging or electrophysiological data. Once trained, these data-constrained RNNs let us infer matrices quantifying the interactions between all pairs of modeled units. Such model-derived “directed interaction matrices” can then be used to separately compute excitatory and inhibitory input currents that drive a given neuron from all other neurons. Therefore different current sources can be de-mixed – either within the same region or from other regions, potentially brain-wide – which collectively give rise to the population dynamics observed experimentally. Source de-mixed currents obtained through CURBD allow an unprecedented view into multi-region mechanisms inaccessible from measurements alone.

We have applied this method successfully to several types of neural data from our experimental collaborators, e.g., zebrafish (Deisseroth lab, Stanford), mice (Harvey lab, Harvard), monkeys (Rudebeck lab, Sinai), and humans (Rutishauser lab, Cedars Sinai), where we have discovered both directed interactions brain wide and inter-area currents during different types of behaviors. With this powerful framework based on data-constrained multi-region RNNs and CURrent Based Decomposition (CURBD), we ask if there are conserved multi-region mechanisms across different species, as well as identify key divergences.

Join Carney’s Center for Computational Brain Science (CCBS) for a seminar on “Practical sample-efficient Bayesian inference for models with and without likelihoods.” This event will feature Luigi Acerbi, Ph.D., an assistant professor at the University of Helsinki.

Abstract:
Bayesian inference in applied fields of science and engineering can be challenging because in the best-case scenario the likelihood is a black-box (e.g., mildly-to-very expensive, no gradients) and more often than not it is not even available, with the researcher being only able to simulate data from the model. In this talk, I review a recent sample-efficient framework for approximate Bayesian inference, Variational Bayesian Monte Carlo (VBMC), which uses only a limited number of potentially noisy log-likelihood evaluations. VBMC produces both a nonparametric approximation of the posterior distribution and an approximate lower bound of the model evidence, useful for model selection. VBMC combines well with a technique we (re)introduced, inverse binomial sampling (IBS), that obtains unbiased and normally-distributed estimates of the log-likelihood via simulation. VBMC has been tested on many real problems (up to 10 dimensions) from computational and cognitive neuroscience, with and without likelihoods. Our method performed consistently well in reconstructing the ground-truth posterior and model evidence with a limited budget of evaluations, showing promise as a general tool for black-box, sample-efficient approximate inference — with exciting potential extensions to more complex cases.

Links:

• MATLAB toolbox
• VBMC paper
• VBMC with noisy likelihoods paper
• Inverse binomial sampling preprint

Join Carney’s Center for Computational Brain Science (CCBS) for a seminar on “Neural reinforcement: re-entering and refining neural dynamics leading to desirable outcomes.” This event will feature Vivek Athalye, Ph.D., postdoctoral researcher at Columbia University.

“Geometry of Object Representation in Visual Hierarchies”

Haim Sompolinsky, Ph.D.
The Hebrew University  

Abstract: Neurons in object representations in top stages of the visual hierarchy exhibit high selectivity to object identity as well as to identity-preserving variables, including location, orientation and scale. suggesting that changes in the object representations from low to high processing stages are related to changes in the geometry of object manifolds. Each manifold consists of the set of population responses to stimuli belonging to the same object.

In my talk, I will present recent work that elucidates the relation between manifold geometry and object-identity computations. I will discuss two kinds of computations. The first is object classification. I will describe new measures of manifold radius and dimensions that predict the ability to support object classification (Chung et al., PRX, 2018). Based on these measures, we characterize the changes in manifold geometry as signals propagate across layers of Deep Convolutional Neural Networks (DCNNs). Recordings from neurons in various stages of the visual systems, have been similarly analyzed, allowing us to test the correspondence between DCNNs and the visual hierarchy in the visual cortex.

In a recent unpublished work with Ben Sorscher (Stanford), we have studied the ability to learn new objects and object categories from just a few examples (the few shot learning problem). We show that feature layers in DCNNs exhibit a remarkable ability in few shot learning of new categories. To explain this performance, we develop a new theory of the geometry of concept formation, that delineates the salient geometric features that underlie rapid concept formation in artificial and brain sensory hierarchies.

Please join Carney’s Center for Computational Brain Science (CCBS) on November 18 for a special seminar on “Differential Resilience of Neurons and Networks with Similar Behavior to Perturbation,” featuring Eve Marder, Ph.D., university professor and Victor and Gwendolyn Beinfield Professor of Biology at Brandeis University.

Please note, you must be logged into Zoom through your Brown account to join this event. 

Abstract:

Both computational and experimental results in single neurons and small networks demonstrate that very similar network function can result from quite disparate sets of neuronal and network parameters. Using the crustacean stomatogastric nervous system, we study the influence of these differences in underlying structure on differential resilience of individuals to a variety of environmental perturbations, including changes in temperature, pH, potassium concentration and neuromodulation. We show that neurons with many different kinds of ion channels can smoothly move through different mechanisms in generating their activity patterns, thus extending their dynamic range.

The Brown DeepLabCut+ Users Group, hosted by the Carney’s Center for Computational Brain Science, will hold its inaugural meeting Thursday, October 8, 2-3 p.m.

The goal of the user group is to provide a local community for peer support and information sharing, and to guide future decisions on local resources, such as support for running DLC on Oscar.

In this first meeting, moderated by Jason Ritt, Carney’s scientific director of quantitative neuroscience, and Maria Daigle, research assistant in the Department of Neuroscience, we will discuss the group’s organization, and invite all attendees to share any issues they are facing using DLC in their research, and/or troubleshooting advice. Going forward we expect to collect and share technical documentation, and provide a forum to match users needing help with local expertise.

Please respond prior to the meeting to this short questionnaire.

“Information and randomization in exploration and exploitation,” featuring:

Robert Wilson, Ph.D.

Assistant Professor

University of Arizona

Abstract: The explore-exploit dilemma is a fundamental behavioral dilemma faced
by any organism that can learn. Should we explore new options in the
hopes of learning something new or exploit options we already know to
be good. In this talk I will present evidence that people use two
distinct strategies for solving the explore-exploit dilemma: directed
exploration, in which information seeking drives exploration by
choice, and random exploration, in which behavioral variability drives
exploration by chance. In addition I will present initial evidence
showing that these two types of exploration rely on dissociable neural
systems.

Contact tiantian_li1@brown.edu for the Zoom meeting password.

Anna Schapiro, Ph.D.

University of Pennsylvania

“Learning distributed representations in the human brain”

 Join via Zoom: https://brown.zoom.us/j/95275026350

Please note, this seminar requires you to be logged into Zoom through your Brown account. Click to learn more.

Abstract: The remarkable success of neural network models in machine learning has relied on the use of distributed representations — activity patterns that overlap across related inputs. Under what conditions does the brain also rely on distributed representations for learning? There are benefits and costs to this form of representation: it allows rapid, efficient learning and generalization, but is highly susceptible to interference. We recently developed a neural network model of the hippocampus that proposes that one subregion (CA1) may employ this form of representation, complementing known pattern-separated representations in other subregions. This provides an exciting domain to test ideas about learning with distributed representations, as the hippocampus learns much more quickly than the neocortical areas that have often been proposed to contain these representations. I will present modeling and empirical work that provide support for the idea that parts of the hippocampus do indeed learn using distributed representations. I will also present ideas about how hippocampal and neocortical areas may interact during sleep to further transform these representations over time.

Nowadays the technology to create videogames and impressive photorealistic scenes is for everyone to reach. Frameworks such as Unity 3D Engine are easy to understand and work with or without any coding experience. So Why not using it to visualize data? You will learn basics on game objects components, UI design and concepts of volume rendering. Just bring your laptop with Unity Engine installed.

GitHub Actions is a relatively new service by GitHub that allows you to run tests, deploy code, and more all from within your repo. This workshop will cover some of the basic use cases of GitHub Actions as well as show some of the more creative ways it can be used to automate your workflows. Attendees should be familiar with Git and GitHub.

Instructor: Mary McGrath 

Intro to PyTorch, by Minju Jung

Nowadays, PyTorch is one of the most popular deep learning libraries. In this tutorial, I will cover the basic operations of PyTorch and building a neural network model for image classification. The basic knowledge of deep learning and coding experience with python are required.

Data Science Computation and Visualization Workshop

EXPLORATORY DATA ANALYSIS WITH PANDAS IN PYTHON, PART TWO with Andras Zsom

Exploratory data analysis (EDA) is the first step of any data science project. In the second part of this pandas tutorial, I’ll walk through various visualization types you can use to better understand the properties of your data at a glance using pandas. Coding experience with python is required but no experience with the pandas package is necessary to follow the tutorial.

Friday 2/28 @ 12pm

Carney Innovation Space, 4th Floor

164 Angell Street

Pizzas and sodas will be served. Sponsored by the Data Science Initiative and organized by the Center for Computation and Visualization.

Data Science Computation and Visualization Workshop

EXPLORATORY DATA ANALYSIS WITH PANDAS IN PYTHON, PART ONE with Andras Zsom

 Exploratory data analysis (EDA) is the first step of any data science project. In the first part of this pandas tutorial, I’ll walk through how to read in csv, excel, and sql data into a pandas data frame, how to select specific rows and columns based on index or condition, and how to merge and append various data frames. Coding experience with python is required but no experience with the pandas package is necessary to follow the tutorial.

Friday 2/14 @ 12pm

Carney Innovation Space, 4th Floor

164 Angell Street

Pizzas and sodas will be served. Sponsored by the Data Science Initiative and organized by the Center for Computation and Visualization.

Data Science Computing and Visualization Workshop (DSCoV)

Topic: DataLad
Instructor: Yaroslav Halchenko

DataLad provides a data portal and a versioning system for everyone, DataLad lets you have your data and control it too. For details, see https://www.datalad.org/.

Friday, November 22 , 12:00 PM
164 Angell Street, 4th Floor Innovation Space
Organized by CCV; Sponsored by DSI