BRAIN Initiative TMM Data Reuse Abstracts
Back to Main Data Reuse page
PARK, IL MEMMING (contact); PILLOW, JONATHAN WILLIAM Real-time statistical algorithms for controlling neural dynamics and behavior EB026946
- Method for inferring latent dynamical system and neural state trajectory from spike trains. Real-time machine learning tool for time series visualization and dimensionality reduction.
- Do low-dimensional continuous trajectories explain the spatiotemporal structure in the neural recordings? If so, what is the underlying dynamical system that governs the neural recording? How are the various task variables related to the neural recording via the low-dimensional manifold?
- High-dimensional neural time series recorded while the animal is engaged in a simple behavior with simple stimulus and a small number of randomized task variables
- Continuous recording in high-sampling frequency (500Hz or higher), sorted spike trains and multi-unit activities. Our method can run in real-time while the recording is happening.
DOIRON, BRENT D (contact); SMITH, MATTHEW A; YU, BYRON M Neuronal population dynamics within and across cortical areas EB026953
A major goal of theoretical neuroscience is to develop mechanistic network models that can reproduce key aspects of neuronal activity recorded in the brain. There are two key parts of fitting a network model to neuronal recordings: 1) incisive measures to compare neuronal recordings with the activity produced by network models, and 2) automatic methods to efficiently fit network model parameters to data. We propose a systematic framework using population activity statistics based on dimensionality reduction and a Bayesian optimization algorithm to efficiently fit model parameters to data. The proposed population statistics go beyond the commonly-used single-neuron and pairwise statistics and raise the bar for comparing models to data. The Bayesian optimization algorithm efficiently fit the parameters using fewer iterations than brute force methods. We emphasize limits of model capacity where a given model reproduces some, but not all, of the desired features of neuronal recordings. We used our algorithm to study which aspects of neuronal activity recorded in macaque V4 can be reproduced by classical balanced networks (CBN) and spatial balanced networks (SBN). We found that SBN has better capacity compared to CBNs in fitting V4 data and discovered interesting trade-offs between different types of activity statistics, thereby revealing limits of model capacity. These insights can be used to guide the development of future network models whose activity resembles neuronal recordings even more closely.
Modeling the dynamics of history dependent neuronal systems at all scales
EB026939
Fidel Santamaria, PI
My lab has developed a mathematical and computational framework to model and analyze history dependent processes, from the diffusion of molecules inside, on the surface, and around neurons, to electrical network activity. What I can offer to experimentalists are unique tools to study scale free processes. This is not your traditional scale free statistical studies that look at long tail probability distributions, instead we can write and model differential equations that give you access to the dynamics of the problem. Our advantage is that we use fractional order integro-differential equations. This mathematical objects are the natural tool to study history dependent phenomena. The type of data I like are long term recordings, either in resting or active states. These recordings can be fluorescent traces from synaptic or neuronal activity, EEG, single- or multi-unit recordings. As an example, my collaborator in this grant, Maurice Chacron at McGill, records from weakly electric fish as they receive natural stimuli. We have been able to replicate the non-linear response of the neurons he records from using models but also, recently, implementing neuromorphic circuits.
HOWARD, MARC W Toward a Theory for Macroscopic Neural Computation Based on Laplace Transform EB022864
We have worked on a theory for how populations of neurons represent information and manipulate information. The theory interfaces well with cognitive models, especially for working memory tasks. The theory predicts that neural populations come in pairs. The optimal data for us has spikes from many simultaneously recorded neurons from more than one brain region during a complex behavior that we can analyze. The model is scale-invariant so very ``slow'' tasks (extended over more than a minute per trial) are of special interest.
LYTTON, WILLIAM W (contact); ANTIC, SRDJAN D Embedded Ensemble Encoding EB022903
https://www.imagwiki.nibib.nih.gov/sites/default/files/2020-11/WikImage_Antic%2BLytton.mp4
We have developed a detailed model of signaling within a rodent Layer 5 neocortical pyramidal cell showing NMDA-mediated plateau potentials lasting 100-400 ms with forward- and back-propagating action potentials. These simulations offer the opportunity to reconceptualize the role of the pyramidal neurons as having complex spatiotemporal dendritic properties with localized UP-states that spread to the soma.
DRUCKMANN, SHAUL Dissecting distributed representations by advanced population activity analysis methods and modeling EB028171
The tool we are developing aims to distill simultaneous recordings from neural populations (e.g., from two brain areas) into a spatio-temporal profile of strength of influence. This is the first year of the TMM grant and accordingly we are very much in the development phase. Our approach defines influence by the ability to predict unexpected deviations in the dynamics of one brain area, the modeled area, from the just-preceding activity in another brain area, the influencing area. In more detail, we first predict the dynamics of the modeled area from its own past history. We then detect deviations from the predicted dynamics and determine whether these deviations can be themselves reliably predicted from the state of the influencing area. There are numerous variants of this general description, such as using linear vs. non-linear predictive models, or modeling the full activity space our inferred subsets. We will complement it with non-dynamical approaches that optimize over both populations to find components of maximal correlation such as canonical correlation analysis Our current use involves electrophysiology data, though we would like to extend our approach in the future to calcium recordings. In terms of data requirements: signal-to-noise ratios vary a lot between tasks and brain areas making it hard to define a specific minimal number of neurons but this method is meant for population recordings, e.g., more than ten neurons per population to be modeled. As it is a sub-single trial method, repetitions are required, e.g., 50 repetitions of a task condition, or extended recordings in a non task-based structure.
YE, BING (contact); DIERSSEN, MARA New methods and theories to interrogate organizational principles from single cell to neuronal networks EB028159
Our project “New Methods and Theories to interrogate Organizational Principles from Single Cell to Neuronal Networks” aims to develop a user-friendly modeling toolset to study how single neurons morphology can determine the connectivity pattern of the network and shed light to the rules linking both. The connectivity patterns of a particular brain region will be estimated by generating a morphological neural network model that uses both neural population data and single neuronal reconstruction data extracted from fluorescent whole brain images. We already developed a population analysis tool to compute the location and orientation of each neuron relative to a reference coordinate system. Our tool is able to overcome the three limitations that are commonly found in cell detection algorithms: undetected neurons, false positives in axonal regions and out of memory errors that arise while processing whole brain images. Our software is Open Source; the source code of our population analysis tool, mainly written in Python, will be available in GitHub. User Friendly, as it can be managed through Graphical User Interface (GUI) or Command Line Interface (CLI). Cross-platform, as it can be executed over Linux and Windows. Big Data oriented: our algorithm is able to compute in reasonable time neuronal location and orientation from whole brain images with a resolution of tenths of microns per voxel and a memory size around the order of tens of Terabytes. Computational Tractability: it is able to split whole brain images into small overlapping 3D-images to avoid runtime out of memory errors. Computational Efficiency: the user can select parallel computing parameters to speed up computing time. Hardware Scalability. The performance is scalable to available computing resources and can be executed on a regular laptop, a workstation or a computer cluster. Vaa3d visualization: the location of detected neurons can be stored in a file format allowing visualization using the rendering power of Vaa3D software. The beta version of the Population Analysis Tool will be released very soon and 1) We are seeking labs interested on using our software to evaluate the usability of the tool and to identify those missing functionalities. 2) We need mesoscopic image datasets taken from different microscopes to evaluate the robustness of the tool and to provide support for more image formats and 3) we are interested in whole brain images with high density neural labeling.
ENGEL, TATIANA Discovering dynamic computations from large-scale neural activity recordings EB026949
Core brain functions—perception, attention, decision-making—emerge from complex patterns of neural activity coordinated within local microcircuits and across brain regions, with dynamics down to milliseconds. Recently, massively-parallel technologies enabled activity recordings from many neurons simultaneously, offering the opportunity to investigate how activity is orchestrated across neural populations to drive behavior. To reveal dynamic features in these large-scale datasets, computational methods are needed that can uncover neural population dynamics and identify how individual neurons contribute to the population activity. Existing methods rely on fitting ad hoc parametric models to data, which often leads to ambiguous model comparisons and estimation biases, limiting the potential of these methods for scientific discovery. To push these limits, our BRAIN project team develops a broadly applicable, non-parametric inference framework for discovering population dynamics directly from the data without a priori model assumptions. Our non-parametric methods explore the entire space of all possible dynamics in search of the model consistent with the data, leading to a conceptual shift from model fitting to model discovery. This is achieved by extending latent dynamical models to a general form, where the latent dynamics are governed by arbitrary dynamical-systems equations, in which driving forces are directly optimized. Our framework reconstructs population dynamics with millisecond precision on single trials and infers idiosyncratic relationships between single-neuron firing-rates and the population dynamics, revealing heterogeneous contributions of single neurons to circuit-level computations. With our methods, we examine large-scale physiological recordings during decision-making, to reveal how neural activity is coordinated to drive decisions and how functional heterogeneity of single-neuron responses aligns with anatomical organization of decision-making circuits.
A python package is available on GitHub: https://github.com/engellab/neuralflow
SOMMER, FRIEDRICH T Building analysis tools and a theory framework for inferring principles of neural computation from multi-scale organization in brain recordings EB026955
Abstract 1:
Title: Identifying correlates of behavior in multi-electrode LFP recordings
Oscillations in the local field potential (LFP) have historically been viewed as coarse-grained indicators of behavioral state. A challenge in understanding the LFP is that it is composed of responses of many thousands of cells and that it is dominated by spontaneous activity, not directly coupled to observable behavioral events or stimuli. Using multi-electrode LFP recordings from the hippocampus, we developed a method to extract precise behavioral information that is embedded within spatio-temporal oscillatory patterns. We have recently extended this approach to extract information from signals that are only weakly and intermittently oscillatory. Not only does our approach offer a robust alternative to spike-based brain-machine interfaces, it suggests how large-scale population codes are embedded within brain dynamics that could subserve inter-regional computation and communication.
Our LFP decoding tool would be most useful for groups interested in identifying the behavioral information embedded within a particular brain region of interest. The data would consist of simultaneous LFP recordings from at least a few dozen sites (the more sites the better), in addition to recordings of relevant behavioral variables (e.g. stimulus properties and behavior). The analysis pipeline offers both supervised and unsupervised modes for identifying dependencies between distributed LFP patterns and behavior. While we have applied this to data sampled in the hippocampus at 25 Hz over ~30 minutes, in principle our approach can be applied to recordings at any sampling rate, as long as there is at least some identifiable oscillatory activity within the signal.
Abstract 2:
Title: Multiplicative encoding of position and head orientation in multichannel hippocampal LFP
Previous work has shown that hippocampal theta-band local field potentials (LFPs) robustly encode position in rats navigating a linear track. This encoding scheme becomes visibly salient by applying ICA to the multi-channel LFP, producing position-tuned components reminiscent of place fields. However, the position tuning is absent in the ICA output of 64-channel LFP recordings of rats foraging in an open field. This is surprising because simulations of place cell-generated LFP predict place-tuning in both the linear track and open field. We hypothesized that this disparity arises from the fact that position is jointly encoded with the rat’s orientation. We explored this hypothesis by analyzing (1) simulated LFP of rats in the open field, containing multiplicatively encoded position/orientation; (2) 256-channel recordings of CA1 LFP from rats in the open field. The results show that jointly position/orientation-tuned components are gradually resolvable as more channels are added. Our simulations and experimental analyses are captured in a polished and well-documented set of Jupyter notebooks (our “tool”) that may be of broad interest to analysts of electrophysiological data.
CARLSON, DAVID E Uncovering Population-Level Cellular Relationships to Behavior via Mesoscale Networks EB026937
The overarching goal of this proposal is to learn how neurons’ action potentials, long considered to be a fundamental unit of information, relate to whole-brain spatiotemporal voltage patterns and behavior. To uncover this relationship, we will develop novel computational methods capable of learning networks that relate voltage signals from multiple brain regions based upon our previously developed explainable machine learning approach. These networks will then be used to stratify neurons into subtypes consistent across a population of subjects to facilitate the statistical aggregation of data to uncover relationships between multiple scale of neural activity and behavior.
What analytical (Theories, Models and Methods) tools have you developed?
We are developing models of mesoscale network activity from implanted, multi-site electrodes to integrate whole brain activity.
What questions can you answer?
Our immediate goal is to use these techniques to gain a greater understanding of how mesoscale networks related to neuropsychiatric disorders and to more basic units of information (e.g., neural firing)
What input do you need? (e.g., cellular activity, sub-cellular, sensory input, complex behavior)
Validation of our methods requires multi-region Local Field Potential recordings (also amenable to EEG measurements), ideally with paired behaviors and neural activity.
CHING, SHINUNG Efficient resource allocation and information retention in working memory circuits EB028154
Short-term working memory is critical for all cognition. It is important to fluid intelligence by definition and is disordered in many psychiatric conditions. It is also an ideal model system for studying the link between the dynamics and functions of neural circuits. Short-term storage requires dynamics that are flexible enough to allow continuous incorporation of new information, yet stable enough to retain information for tens of seconds. Much is known about the neuronal substrate of short-term memory. There is a gap, however, in our knowledge of how neuronal resources are efficiently allocated to store multiple items. This gap is particularly striking given that a multi-item memory task (memory span task) is often used to measure fluid intelligence. Neurons in frontal areas are active during a memory period, and individual neurons are tuned to respond to particular memoranda. It is known that individual cells ramp up or down during a memory period. However, we were surprised to discover in preliminary experiments that 80% of individual cells in memory circuits lose their tuning before the end of a 15s memory period. This loss of tuning occurs at similar times across repeated trials; a neuron that loses tuning at 3s in one trial seldom remains tuned for more than 7s in a subsequent trial, and vice versa. This leads to the question of whether cells with common “drop-out” times are linked together in a subnetwork, similar to the “slot” organization often posited to support multi-item memory. We formulated a theory about how these subnetworks might be organized to enact a form of efficient resource allocation that balances demand for memory capacity against memory duration. The primary goal of our TMM project is to test the validity of this theory, and more generally probe memory circuits for evidence of functional subnetworks, using a unique combination of long-delay multi-item memory tasks, computational modeling and analysis. Our project integrates experimental and computational methods, including formalisms from information and control theories, so as to build tight links between (i) the observed phenomenology; (ii) the mathematical consistency of the theory; and (iii) how (i) and (ii) might be reconciled mechanistically in the dynamics of neural circuits
What analytical (Theories, Models and Methods) tools have you developed?
We are developing bottom-up and top-down circuit-level models to derive new mechanistic understanding of how working memory is encoded. These models are constrained by neuronal biophysics but optimized in order to meet hypothetical high-level functional objectives associated with working memory function.
What questions can you answer?
Our immediate goal is to gain a deeper understanding of how memory resources are encoded and allocated/managed within neural circuits. More broadly, our goal is to develop a general modeling paradigm that can associate dynamics to higher-level circuit function.
What input do you need? (e.g., cellular activity, sub-cellular, sensory input, complex behavior)
Validation of our theory requires recordings of cellular activity alongside behavioral characterizations (working memory performance). Our immediate goals are in the domain of spatial working memory, but it would be of interest to broaden the scope to other memory domains.
What are the data specifications needed for your TMM tool? (e.g. data type, sampling frequency, species type, brain area, modality, cell type, duration of recording)
Our plans are to validate our theory in NHPs with recordings from dorsolateral prefrontal cortex and frontal eye fields while animas are engaged in a spatial working memory task.
DAVID, STEPHEN V Tools for modeling state-dependent sensory encoding by neural populations across spatial and temporal scales EB028155
Models of the functional relationship between dynamic sensory stimuli and neural activity form a foundation of research in sensory neuroscience. The advent of modern machine learning methods has introduced the possibility of new and more powerful models of sensory coding. Studies using convolutional neural networks (CNNs) and related models have shown that they can outperform traditional encoding models, in some cases by a substantial degree. In addition to standard applications describing feed-forward coding by single neurons, these methods can be adapted to multi-channel neural data and to characterization of behavior state-dependent changes in coding. While potentially powerful, CNNs can be challenging to implement and interpret, especially without expertise in computational methods. We have developed the Neural Encoding Model System (NEMS) as a python-based toolkit for fitting both traditional and machine learning models to sensory neurophysiology data.
NEMS was developed for use in the auditory system but it can be applied to any system representing information about dynamic extrinsic signals. It employs a modular design that allows elements from traditional encoding models (linear filters, synaptic plasticity, gain control) to be incorporated into artificial neural network models with broad flexibility fitting algorithms. Models can be fit using either scipy- or Tensorflow-based backends. A scripting system allows scaling to large datasets and compute clusters. The system also streamlines direct, quantitative comparison of a large family of models on the same dataset and characterizing functional equivalence of different model architectures.
Data types: calcium imaging, single unit, EEG
Sampling frequency: 0.1 Hz to 1 kHz
Time scale: 10s of seconds to hours
Modality/area: auditory system, can be adapted to sensory and motor systems.
KRAMER, MARK ALAN (contact); EDEN, URI TZVI Measuring, Modeling, and Modulating Cross-Frequency Coupling EB026938
What is being modeled?: Interactions between different frequency brain rhythms.
Description & purpose of resource: We provide a statistical modeling framework to estimate high frequency amplitude as a function of both the low frequency amplitude and low frequency phase; the result is a measure of phase-amplitude coupling that accounts for changes in the low frequency amplitude. The proposed method successfully detects cross-frequency coupling (CFC) between the low frequency phase or amplitude and the high frequency amplitude, and outperforms an existing method in biologically-motivated examples.
Spatial scales: tissue
Temporal scales: 10-3 - 1 s and 1 - 103 s
This resource is currently: mature and useful in ongoing research
Has this resource been validated?: No
How has the resource been validated?: Details, simulation results, and applications to in vivo data are published.
Key publications (e.g. describing or using resource): A statistical framework to assess cross-frequency coupling while accounting for confounding analysis effects, Nadalin et al eLife 2019;8:e44287
https://elifesciences.org/articles/44287
DOI link to publication describing this resource: https://doi.org/10.7554/eLife.44287
Links: All code to use and further develop this method is available on GitHub.
Keywords: BRAIN, TMM
MAKSE, HERNAN; HOLODNY, ANDREI I Application of the principle of symmetry to neural circuitry: from building blocks to neural synchronization in the connectome EB022720
Grant Title: Application of the principle of symmetry to neural circuitry:
From building blocks to neural synchronization in the connectome
By: Hernan Makse and Manuel Zimmer.
1. Our analytical tool:
We have developed a network theoretical toolbox to extract the
symmetries of the connectome. The symmetries are graph automorphisms
or symmetry permutations, i.e. specific similarities in the
connectivity patterns of the connectome, that predict synchronization
of neural populations. The theory successfully predicted functional
building blocks in the C. elegans connectome, like circuits governing
locomotion (see: https://www.nature.com/articles/s41467-019-12675-8).
2. Questions to answer:
Our central hypothesis to test is if the symmetries in connectivity
underly the synchronization of neuronal population activity, and
therefore can be used to discover functional units within complex
connectome data. Our graph theoretical toolbox will classify the
symmetries of all connectomes thereby identifying neural circuits that
potentially form functional building blocks. Using our symmetry
finder, we aim at predicting which neurons synchronize their activity
and then to further test and investigate these structure-function
relations by (I) measuring neuronal activity and (II) manipulating the
underlying circuits experimentally.
3. What input do you need?
We are calling all connectomes. The archetypical example is the
complete reconstruction of the C. elegans connectome. Partial
reconstructions with similar level of resolution at the neuron-level
and full connectivity are also needed.
4. Specifications:
4.1. Anatomical data: Connectomes could include larval zebrafish,
larval annelid Platynereis, partial reconstructions of the drosophila
adult and larval brain (e.g. visual system or mushroom body) or
partial reconstructions of rodent brains.
4.2. Dynamical data: Alongside these anatomical data, we look for
dynamical single-cell resolution neuronal activity data that can be
acquired from these models, e.g. population wide calcium imaging data
and multi-unit electrophysiological recordings.
MISHNE, GAL Data-driven analysis for neuronal dynamic modeling EB026936
Summary: We are developing methodology for analysis of large-scale neural data, primarily two-photon calcium imaging data. We aim to develop an end-to-end modeling and analysis framework for multi-trial neuronal activity across multiple modalities and spatiotemporal scales:
- low-level processing of raw calcium imaging data
- mid-level organization of extracted interconnected neuronal time-traces
- high-level analysis of evolving network of neurons and behavior over long term learning.
For imaging analysis, we have developed methods for ROI extraction and time-trace demixing of 2p data and parcellation of widefield calcium imaging data. Our approach does not directly model calcium dynamics and can apply to motion-corrected high-dimensional imaging data. To further validate and extend our approaches we would be happy to receive cellular-level 1p data, voltage imaging and spatial transcriptomics datasets.
We are developing tools for visualization of dynamics and analysis of a network of neurons as it learns a task (longitudinal studies), with preliminary results in artificial networks. To apply this tool we need cellular activity from a large identified group of neurons during learning, alongside behavior and/or stimulus.
Finally we are developing tools for unsupervised and semi-supervised behavior annotation. Complex recurring behavior from multiple animals (of the same species) would assist in further developing and validating our approach. For the semi-supervised approach, labels for part of the data / spatial trajectories (such as Deep lab cut) are required.
WITTEN, DANIELA Models and Methods for Calcium Imaging Data with Application to the Allen Brain Observatory EB026908
Abstract for Data-Modeling Match: We have developed a statistical model, algorithm, and corresponding software to estimate the times at which a neuron spiked on the basis of calcium imaging data. We are also in the process of developing a model and algorithm and software to perform inference on these estimated spike times, i.e. to obtain a p-value quantifying how likely it is to have observed such a large increase in fluorescence in the absence of a true spike. Our software is implemented in python and R and is described here: https://jewellsean.github.io/fast-spike-deconvolution/.
What is the analytical tool you have developed: We have developed a model for spike estimation on the basis of calcium imaging data, and are currently developing an approach to conduct inference on the estimated spikes. The model is implemented in R and python, and is available at https://jewellsean.github.io/fast-spike-deconvolution/.
What input do you need?: These methods require calcium imaging data as input. For each neuron, the fluorescence trace should be DF/F transformed.
What are the questions you can answer?: On the basis of calcium imaging data, we can answer the question of "When did the neuron spike?" We are also working to answer the question of "What is the probability of observing such a large increase in fluorescence in the absence of a spike?" (The latter amounts to computing a p-value associated with each spike.)
What are the data specifications needed for your TMM tool?: The TMM tool requires DF/F derived from calcium imaging data, for a single neuron. It can be applied repeatedly to a large number of neurons. So far we have applied it to data from the visual cortex of mice, from the Allen Brain Observatory. We are in the process of applying it to dopamine neurons.
SHOUVAL, HAREL ZEEV Learning spatio-temporal statistics from the environment in recurrent networks EB022891
My lab is interested in how circuits of neurons can learn and represent the spatio-temporal dynamics of external stimuli. We have developed models with spiking neurons and local biophysically plausible learning rules that can accomplish this. In our framework the ability of circuits to accomplish this task depends a pre-existing local structure of cortical microcircuits. What we can offer experimentalists is a theoretical framework that can make sense of specific microcircuits circuits in brain, and make sense of specific synaptic plasticity observed experimentally in the brain. We make specific predictions about different classes of temporal profiles of cortical cells. We are interested in both electrophysiological recording results and calcium imaging results both from in vivo experiments and from slice experiments. These should come from experiments in which animals or slices were exposed to patterns or paradigms with temporal regularity over expended time periods of at least hundreds of milliseconds. What we can offer experimental labs is to use unsupervised dimensionality reduction and clustering methods in order to classify single cells within the networks, and correlation methods to uncover effective connectivity. We can use these results to help verify or reject our current models, and to generate hypothesis as to the circuit wide implications of the results. We are also interested in results related to synaptic plasticity in similar types of experiments, and especially in neuromodulator dependent synaptic plasticity. Here too we can analyze the data and offer model-tested hypotheses as to the implications of the experimental results for learning in circuits.
RAJAN, KANAKA Multi-region Network of Networks Recurrent Neural Network Models of Adaptive and Maladaptive Learning EB028166
We in the Rajanlab at Mount Sinai design neural network models constrained by experimental data, and reverse engineer them to figure out how brain circuits function in health and disease. Here are two scalable, flexible, and robust tools we have developed, through our TMM grant (R01EB028166-01), that will find wide adoption across to the broader neuroscience community, particularly in a few U19 consortia. We are already collaborating fruitfully with the BRAIN_COGS consortium at Princeton University and would be delighted to share our collective findings at the upcoming meeting. That said, we are always looking for opportunities to collaborate with experimental labs and are particularly keen to get involved at the ground-floor of such collaborations, i.e., having a key role in designing experiments, not only leveraging existing data.
1) The first tool we have developed is named Current-based Decomposition or CURBD for short. This powerful new theory-based framework is intended for “in-vivo tract tracing” from multi-regional neural activity collected experimentally. 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.
2) We have additionally developed a second data-inspired recurrent neural network (RNN)-based tool, termed TRAKR. This tool enables us to detect state transitions from time series data. Using TRAKR, we can identify neural state transitions that are predictive of animal behavior under different experimental and experiential conditions. For example, we can pick up behavioral, neural, and circuit mechanistic changes in decision making, those that occur during shifts in attention or task engagement in multi-task paradigms, those accompanying learning of such tasks in the lab, etc. The current implementation of TRAKR works with time series data such as behavioral monitoring from pose detection methodologies, EEGs, ECoGs, LFPs, spiking neural data, and calcium fluorescence signals. This is currently a human-in-the-loop method that detects state transitions after correlation with behavior, so behavioral recordings are helpful but not necessary. There is no threshold for sampling frequency, but larger datasets are desired since RNNs are more robust when supplied with greater amounts of data. This tool is agnostic to the species, brain areas, or cell types. It is similarly oblivious to modality and can work equally well with large amounts of EEG, LFP, and even imaging time series.
KILPATRICK, ZACHARY PETER Connecting neural circuit architecture and experience-driven probabilistic computations EB029847
What is the analytical tool you have developed?
Theory/Models (for now): Specifically, we are developing mechanistic models that account for large-scale activity in neural circuits responsible for both persistent and dynamic activity states during the delay period of working memory tasks.
What input do you need?
At this stage, we are primarily looking to constrain our models using psychophysical data from human behavioral experiments generated by our collaborators (Gold: Penn; Buschman: Princeton; Bays: Cambdrige). At a later stage, we will extend our models to consider more complex network architectures, and at that stage will leverage large scale neural recordings from non-human primates. We are open to any additional response or neural recording data from delayed estimation/oculomotor delayed- response tasks that groups might have to offer that would be relevant to the aims of this project.
What are the questions you can answer?
1) How are experience-driven probabilistic priors represented in neural circuits?
2) How do the dynamics and structure of network architecture contribute to the dynamics of neural activity during the delay period of working memory tasks?
3) How do the representations of multiple stimuli interact across time in neural circuits participating in delayed response tasks?
What are the data specifications needed for your TMM tool?
Data type: Behavioral responses from parametric delayed response tasks (recalling color, orientation, location after a few seconds): stimulus parameters on each trial, delay time, response accuracy/location, time between trials. Pupillometry data would also be helpful (eye tracking).
Neural recordings (multichannel LFP or spike) from working memory relevant areas during the entire task (fixation, delay, response, intertrial).
Sampling frequency: Neural recording sampling frequency at or above 250Hz preferred. Sampling on all trials for behavioral data is preferred.
Species type: Humans and non-human primates would be best.
Brain area: prefrontal, cingulate, parietal, premotor cortices
Cell types: Pyramidal and interneurons