Konrad P. Kording - US grants

Abstract In the proposed research, we will characterize how the nervous system deals with uncertainty in motor learning. Subjects will move a cursor from a starting position to a target position in a virtual environment. Visual feedback will be manipulated to induce uncertainty about the state, the feedback, or its relevance. Our experiments will focus on probing the resulting trial-by-trial learning. The proposed analysis of the influence of uncertainty on motor learning is driven by strong hypotheses derived from a statistical framework. With the expected results we will either be able to refute Bayesian models that formalize how uncertainty affects learning or refute state space models that assume that uncertainty has no influence on learning. Importantly,uncertainty is a central factor for human behavior and quantitatively understanding its role is important beyond any specific modeling framework. The long term objectives of this research program are to answer basic and important questions in motor learning from a computational perspective and to provide tools for improving motor rehabilitation. The nervous system needs to learn in the presence of uncertainty within the functions of everyday life, and in the presence of disease. Based on statistical insights, this study will test key factors that affect the way the nervous system learns from visuo-motor errors. Specifically, we will understand how the times, magnitudes and the visual presentation of errors affect motor learning. Choices in robotic rehabilitation approaches result in how error feedback can be made effective and relevant through maximizing research testing. As we ask fundamental questions, the results are expected to generalize to a wide range of motor learning tasks.

Core B: Bayesian and decision theoretic tools The production of movement sequences is inherently affected by uncertainty: to move rapidly the animal needs to estimate what to do next given past knowledge. Such estimates can never be certain. A colorful example from a recently popular book (Taleb, 2008) shows that we can never be certain about a sequence of events. The turkey that has been fed every day for close to a year gets slaughtered for Thanksgiving. Many communities such as robotics, economics, data mining and models of human behavior are converging on a common approach towards formalizing uncertainty: Bayesian decision theory. We will first use these methods to predict behaviors from each of the three experimental labs. We will continue to extract the relevant variables (timescales, probabilities) that need to be represented by the nervous system to efficiently produce sequences. These variables will then be correlated with measured neural signals to ask how these variables are represented. Moreover, uncertainty is central when analyzing data from neurons. When we are asking how neurons store and recall motor sequences we never directly measure the relevant variables, such as niemory, we rather measure spikes or imaging signals that are affected by noise. A central topic for neural data analysis, therefore, is to combine many measurements (say 1000 spikes) into an estimate (of say tuning properties) that has small uncertainty (or narrow error-bars). We will use state of the art Bayesian data analysis techniques to analyze the data resulting from the proposed experiments in the other projects. Specifically we are interested in asking how neurons interact with one another using these Bayesian methods. Lastly, we will use state of the art decoding methods to ask how well various types of information are encoded by the measured signals. This is useful for the experimental projects as it allows asking how much information about a, variable of interest is encoded by neural signals. RELEVANCE (See instructions): The proposed work is central to the problem of understanding the mechansims where practice leads to to reorganizafion of the human motor system in the face of aging, neurodenerafion, stroke or brain injury. Understanding these mechansims has an impact on the design of therapies directed at preserving function, developing compensator movements and ulfimately, developing novel motor capacity.

CPS: Large: Cybernetic interfaces for the restoration of human movement through functional electrical stimulation

The objective of this research is to develop an intuitive user interface for functional electrical stimulation (FES), which uses surgically-implanted electrodes to stimulate muscles in spinal cord-injured (SCI) patients. The challenge is to enable high-level tetraplegic patients to regain the use of their own arm. The approach is to develop a multi-modal Bayesian user-intent decoder; use natural muscle synergies to generate appropriate low-dimensional muscle activation signals in a feedforward controller; develop a feedback controller to enhance the performance of the feedforward controller; and test the system with SCI patients on daily living tasks, such as reaching, grasping, and eating.
The challenge problem of restoring arm use to SCI patients will lead to new design principles for cyber-physical systems interfacing neural and biological systems with engineered computation and electrical power systems. The tight integration of the proposed user interface and controller with the users own control system requires a deep understanding of biological design principles such as nested feedback loops at different time and length scales, noisy signals, parallel processing, and highly coupled neuromechanical systems.
This work will lead to new technology that dramatically improves the lives of spinal cord-injured patients. These patients often have no cognitive impairment and have long life spans after injury. The goal is to enable these patients to eat, reach, and grasp nearby objects. These tasks are critical for independent living and quality of life. This work will also help train a new generation of students in human-machine interfaces at the undergraduate, graduate, and postdoctoral levels.

Experiments in which subjects reach from one point to another are one of the workhorses of movement science. Many questions ranging from basic neuroscience (How do neurons represent movement?) to clinical (How do patients with Parkinson's Disease move differently from healthy controls?) to behavioral (How does loud noise lead to movement errors?) are studied using this paradigm. Moreover, many models have been constructed to describe reaching, and these models have strong links to robotics and computer science. This project will develop a database that contains both experimental results as well as models of reaching movement. The database will make it easier for experiments to falsify models and for models to be designed so that they overcome the limitations of previous models.

The objective of the joint database design is that multiple models should be able to make predictions for a given experimental dataset, and likewise, multiple experimental datasets can be used to constrain a given model. The project will start with a relatively narrow set of experiments and models, widening the scope gradually over the course of the project. Research on reaching spans a broad set of questions and a broad set of experimental methods; however, many experimental approaches share significant aspects. The objective of the project is to enable inclusion of the results of a broad set of communities while keeping the database sufficiently coherent to be useful.

To enable the inclusion of a broad set of participants who will share models and data the proposed project will include summer schools, workshops and competitions where participants can compare models. Access to the database will be free and open, and this possibility of access to high quality data promises to allow scientists from any movement related discipline to productively interact with movement data and models.

DESCRIPTION (provided by applicant): There is a fundamental gap in our understanding of how uncertainty is represented during reaching movements. Our ability to see our hand and the potential targets of our reach constantly changes: only when we foveate a well illuminated object can we know precisely where it is; in general there is broadly varying uncertainty about relative location of hand and target. The issue of how neurons encode such kinds of uncertainty for sensory and motor tasks is possibly the most active research area in computational neuroscience and a workshop on this topic organized by Lengyel and others at the 2010 Cosyne meeting was attended by more than 100 scientists. The long term goal of the proposed research is to understand how the nervous system integrates information during reaching and to use this knowledge to accelerate recovery from neuromotor diseases. The objective of this particular application is to quantify how uncertainty affects neural activities in the sensorimotor pathway. The central hypothesis is that one of the theoretically proposed models or a combination of these models will account for the neural representation of uncertainty. The rationale for the proposed research is that a better understanding of the way the nervous system represents uncertainty and promises to improve rehabilitation from neuromotor diseases because uncertainty has been shown to modulate learning speeds. Guided by preliminary data that shows our ability to perform the proposed research by pursuing three specific aims: 1) We will analyze how uncertainty about the general situation, acquired over time and called prior is represented. 2) We will analyze how uncertainty about the current feedback, called likelihood is represented. 3) We will analyze how the nervous system learns about uncertainty. The approach is innovative because it utilizes a highly integrated approach to neuroscience where advanced modeling and new data analysis is directly integrated with experiment design. The proposed research is significant, because it is expected to vertically advance our understanding of the representation of uncertainty and allows a clear distinction between competing and widely held hypotheses. Ultimately, such knowledge has the potential to inform the development of physical rehabilitation therapies. Since sensorimotor uncertainty increases as we age and with a wide range of diseases, a better understanding of the neural basis of uncertainty promises to help reduce the growing problems of an aging population.

The overall goal of these experiments is to understand how the brain controls where we look. To accomplish this, it is important to study brain activity and behavior under conditions that closely approximate those in the real world. All of the experiments we propose to do will use awake behaving rhesus monkeys as subjects. In prior work, we have studied activity in the cortical frontal eye field while monkeys looked at images of natural scenes. The frontal eye field (FEF) is closely involved in the control of purposive voluntary eye movements. While the monkey searched for a target hidden in the images of natural scenes, the activity of FEF neurons consisted of combinations of activity related to planning upcoming eye movements, as well as activity that was sensitive to salient visual features of the image. In parallel with the development of our understanding of how the brain controls eye movements, there have been substantial advances in our understanding of the features of natural images that guide both human and monkey eye movements. These behavioral studies are at the advanced level of being able to accurately predict patterns of eye movements. Our goal in this proposal is to take advantage of these advancements in predicting patterns of eye movements in natural environments to help us understand the brain events that are responsible for this behavior. We will focus upon neuron activity in the FEF due to its essential role in the control of voluntary eye movements. The proposal has 3 Aims each focused upon a different factor that is known to guide eye movements under natural conditions. Salience describes how different a small part of a visual scene is from the remainder of the scene based upon stimulus features such as color, contrast, shape, and orientation. Our first aim will define the effects that salience has upon FEF activity. In our second aim, we'll quantify the effects of relevance. Relevance refers to the importance of visual features for the task at hand; for example, if we're looking for a red target, the red items in the image will be more likely to attract our attention and ultimately be the target for an eye movement. Knowing the broad composition of a scene, a quality that is called scene gist, can tell us the places where an object is more likely to be found. For example, if we are looking for a bicycle, we are more likely to search the sidewalks and roadways of a street scene and ignore other places where bicycles are unlikely to be found. Our final aim will look for the effects of scene gist upon monkey behavior and the FEF activity driving that behavior. In addition to the brain recording experiments outlined above, a large part of our effort will be devoted to mathematical analysis and modeling of the behavioral and neuronal data we obtain. Our ultimate goal is to provide a model that predicts the contributions of salience, relevance, and gist to the activity of FEF neurons. The successful model will be a mathematical representation that predicts search-related activity in the FEF for both artificial and real world conditions.

DESCRIPTION (provided by applicant): The proposed symposium entitled Computational and Translational Motor Control (CTMC) will happen in the autumn of 2013. Its objective is to facilitate the information transfer between scientific communities studying computational motor control (mathematical movement science) and those scientists working on making an impact on translational motor control (bench to bedside). This is important for the mission of the NIH, as it enables research that promises to improve the lives of many patients suffering from neuromotor diseases.

PI: Sensinger, J. W.; Hargrove, L.; and Kording, K. P.
Proposal Number: 1317379

Problem Description: Better robotic prostheses can dramatically improve the quality of life for the more than 40,000 Americans with an upper limb amputation, many of whom reject existing devices because they have trouble controlling them in the same intuitive, subconscious way that they controlled their intact arms. Prosthesis control is difficult because amputees experience great uncertainty both with respect to whether their device will respond appropriately to their control signals and whether sensory feedback cues accurately reflect the actual movement. Researchers have focused on improving isolated aspects of control, for example by improving filters or mimicking able-bodied sensory cues through haptic devices, but these approaches have minimally reduced the uncertainty of prosthesis control. Human interaction with a prosthesis is a multifaceted, time-varying problem that is difficult to solve. What is missing from robotic prosthesis research are principled methods for optimizing control strategies and sensory cues which take into account behavioral choices people are known to make in the face of high uncertainty.

Intellectual Merit: The proposed research is innovative because it poses the co-robot problem in a broader context that incorporates the highly sophisticated behavioral decisions that humans make in optimizing their control strategy and sensory cues. This principled approach is able to integrate multiple effects in ways that were not possible using previous approaches. For example, the proposed approach naturally incorporates the fact that people prefer to use less exerted effort to accomplish a task, but tolerate more effort during portions of movement that require greater precision (e.g. final portion of a trajectory). On the other hand, the approach does not favor high-certainty haptic cues if those cues provide redundant information to existing sensory cues such as vision, or if the haptic information does not reduce the uncertainty of controllable system dynamics. Due to the large sources of control-signal noise present in amputees, the proposed work will lead to improved techniques within the fields of computational motor control and optimal control. This research builds on the team?s extensive experience in the design and control of upper-limb prostheses and in developing the field of computational motor control. Achievement of the proposed aims will contribute to the field of robotic control and to such diverse fields as human-robot interaction, perception, manipulation, and exoskeletons.

Broader Impacts: True biomimetic prostheses, exoskeletons, and humanoid robot control will not be possible until there is a firm understanding of how humans integrate with these co-robots in the face of interacting sources of uncertainty. This computational motor project will provide transformative insight into how humans control movement in the presence of large uncertainty and thus fill a critical gap in the knowledge base of this field. The framework developed in this research will be of great interest to the motor-control research community and may be useful in the restoration of other movement disorders such as spinal cord injury and stroke. The lead institution of this proposal, the Rehabilitation Institute of Chicago (RIC), is consistently ranked the top rehabilitation hospital in the country. The close proximity of research and clinical excellence within RIC ensures that the benefits resulting from this work will be quickly disseminated to prosthesis users. The research team will also seek to reach a broader audience?the laboratories at the RIC are regularly visited by students from local high schools and universities, and the RIC also contributes to outreach activities within inner-city Chicago. These outreach programs promote an awareness of rehabilitation research and an enthusiasm for pursuing a career in engineering. Additionally, the team will develop a K-12 educational module based on the template of the successful Summer School in Computational Sensory-Motor Neuroscience developed at Northwestern University and Queen?s University, which will provide a combination of theory and student-driven experimentation using games that will address many of the Illinois Learning Standards in science, math, and English language arts.

DESCRIPTION (provided by applicant): Progress in neural recording is critical to understanding the brain and developing treatments for brain disorders. Current neural recordings can, at best, capture a few hundred interacting neurons. The number of recorded neurons is relatively small because current neural recording devices, such as electrodes, amplifiers, lasers, and cameras, are macroscopic. The objective of our research is to create neural recorders at the molecular scale, by writing neural activities onto DNA, like a molecular ticker tape. The device will consist of an engineered DNAP polymerase that can be cheaply synthesized and easily delivered to neurons, where it will write the temporal dynamics of activity of each neuron onto local DNA molecules, which can later be analyzed via increasingly cheap genome sequencing technologies. The long term goal of our research is to enable a paradigm shift, making recording instrumentation-free, easy to use, and scalable to arbitrary numbers of neurons. We will obtain the nanoscale recording device using three pipelines: (1) Polymerase design pipeline. We will search through different DNA polymerases to find a polymerase that makes many replication mistakes when ion concentrations increase, and thus when neurons are active. We will use directed protein engineering to add ion-sensitive domains. Lastly we will use high throughput protein directed evolution, to produce a polymerase with desirable properties. (2) Template design pipeline. We will design and deliver an engineered DNA template to the cell to be copied. We will utilize transfection, which is feasible but might not be convenient in some neuroscientific experiments, moving later towards viral template delivery methods, which may be simpler. (3) Statistics pipeline. The resulting DNA sequences need to be converted back into signals of neurobiological meaning. Such conversion needs to be precise, robust to various problems such as biological polymerase noise, and error-correcting. The approach is innovative, because it reinvents the concept of recording using molecular engineering to produce a device that is orders of magnitude smaller and arguably more versatile than comparable devices. The proposed research is significant, because it allows a whole range of new electrophysiological experiments. The approach will complement other emerging approaches that promise to lead to large dataset based neuroscience, e.g. connectomics. The resulting technique will be easyto- use and inexpensive, yet will promise to allow recording simultaneously from potentially arbitrary numbers of neurons, with temporal precision comparable to existing state-of-the-art calcium imaging. It promises massively increased amounts of neural data and entirely new approaches to asking deep questions about the way the brain works and how to cure disease of the brain.

Human telepresence underwater is essential for tasks such as security sweeps in harbors and oil field servicing. Co-robotic solutions are needed, because the risks are great for human divers, while autonomous robots do not deal well with contingencies. A major problem is that vision works poorly in murky environments, such as when mud is kicked up from the bottom. In this National Robotics Initiative (NRI) project the researchers are investigating and developing a replacement for vision -- electrosense -- used by Amazonian fish that navigate and hunt in murky water. These "weakly electric fish" generate an AC electric field that is perturbed by objects nearby. Electroreceptors covering the body of the fish detect the perturbations, which the fish decodes into information about its surroundings. The researchers are developing methods of preprocessing electric images for human understanding, and new computed methods for machine interpretation.

The research creates electrosense hardware and practical testbeds, for navigation and for manipulation underwater. It investigates methods and software to facilitate human interpretation of electric images, as well as machine interpretation. In hardware, the researchers are creating a kilopixel-scale electrosense array as an input sensor for human interpretation of electric images, and development of preprocessing algorithms to make human interpretation workable. The researchers are also using sparser and non-coplanar groups of electroreceptors on a manipulator, for control of pre-grasp and manipulation tasks. For human interpretation, electric image preprocessing includes contour painting and spatial high-pass filtering, as well as temporal filtering. For machine interpretation, methods include specific recognition strategies for simple geometric primitives, and sparse beamforming techniques for more complex environments.

DESCRIPTION (provided by applicant): Current scaling behavior in electrical recording is dominated by the difficulty of fabricating systems with many high bandwidth channels. The objective of our research is a radically new and innovative approach of fabricating massive scale electrical recording setups. In our approach, a polished bundle of insulated metal wires (that act as recording electrodes) are pushed mechanically on the surface of a commercial amplifier chip used for high-speed infrared imaging. This allows us to tie into the massive progress happening in the imaging field. The long term goal of our research is to enable a paradigm shift, making the recording of massive amounts of neurons a cheap possibility. Our approach combines electrode design, innovative methods for electrical connections, and off-the-shelf read out integrated circuits (ROICs). The approach is innovative, because it uses a unique combination of techniques to produce a device that allows orders of magnitude more channels to be recorded at a fraction of the cost with unrivalled potential for future growth. The proposed research is significant, because it is a radical departure from current ways of conducting electrophysiological experiments. The far larger numbers of electrical channels promise to enable a broad range of new experiments, ever boosted by future improvements in imaging chip development. But above all, the planned methods should become compatible with recording from human subjects in the context of brain machine interfaces.

? DESCRIPTION (provided by applicant): Movement rehabilitation is slowly transitioning towards an evidence and theory driven mode of operating where successful treatments are guided by an understanding of how the brain and the body work. As part of this transition, the development of models is very important. In the space of models, Bayesian and optimal control models have been particularly influential over the last decade or two. Models about statistically efficient processing in the brain (Bayesian brain) have been influential and assume that the brain is hardwired for solving statistical problems. Other, generally less popular, models assume that the brain just learns to act in a statistically efficient way from trial and error. For the interpretation of behavioral results this difference is crucial. An answer to this question promise to inform theories, electrophysiology and rehabilitation. Here we propose to use experiments with human subjects and model building to quantify the role of learning for the processing of uncertainty. We compare over trained tasks with novel tasks to ask if humans are hardwired for statistically efficient integration. We compare adults with children to ask if the appropriate experience solving these tasks is necessary. We perform generalization experiments to analyze the neural representation of uncertainty. And lastly, we build models of general learning systems for comparison with existing Bayesian models and test them on large databases. Across the experiments we will juxtapose two-alternative-forced-choice paradigms that quantify uncertainty with sensorimotor integration experiments that allow estimating its effect on behavior. The planned experiments will provide a nuanced answer to the question of how human behavior becomes so efficient in a statistical sense for important tasks. The three aims to be investigated in this project are: Aim 1: Determine if the brain needs to learn how to integrate uncertain sensorimotor information. We will ask if learning is necessary to allow humans to efficiently combine multiple pieces of uncertain information. While many current theories implicitly assume that efficient integration requires no learning our preliminary data suggests otherwise. Studying uncertainty is important as it is ubiquitous and affects learning. Aim 2: Determine if the representation of probability distributions and thus generalization is hardwired. We will ask if th generalization of uncertainty is hardwired or can be learned. This is important because generalization across tasks is one of the prime objectives of movement rehabilitation. Aim 3: Construct and test new models that learn how to integrate uncertain information. We will construct models, based on the idea of deep learning, that can explain statistically efficient processing in the brain without requiring it to be built in. The models will be tested against hardwired Bayesian models and a large database of behaviors. By building strong models that make accurate predictions about current and potential behavior would allow for more efficient and targeted movement rehabilitation.

? DESCRIPTION (provided by applicant): Anatomy defines the `reference' atlas for all of neuroscience. It is one of the most important markers of disease or damage to the brain, and constrains the circuitry of neural computation. However, brain maps are fundamentally incomplete. There remains an information and resolution gap at the mesoscale: whole brain maps visualized at micrometer resolution so that cell shapes, numbers, and positions along with their long range projections can be visualized in a single brain. Such maps, especially if they were compatible with the other modalities, could provide valuable information for determining the cellular composition of brains and serve as a vital bridge between studies using disparate resolutions (e.g. functional magnetic resonance imaging or MRI, and serial section electron microscopy) to image the brain. X-ray microtomography is the only approach that can provide mesoscale detail on whole brains without the need for slicing. We propose to use the nation-leading capabilities of the Advanced Photon Source (APS) at Argonne National Laboratory, which offers a source brightness millions of times higher than laboratory sources. With this source, we have already demonstrated that propagation-based monochromatic phase contrast x-ray imaging can be used to obtain high contrast tomograms of millimeter-sized regions of plastic embedded and metal-stained mouse brain, with data collection times of about a few minutes and a voxel resolution of one micrometer. We will develop whole brain sample preparation methods optimized for x-ray microtomography, and for correlative studies with serial section electron microscopy for synapse-level resolution of small regions (Aim 1). We will develop mosaic x-ray tomography to move from millimeter sized samples to whole mouse brains, with a path towards future studies of human brains (Aim 2). We will develop a high speed tomographic reconstruction workflow, and methods for volume segmentation, analysis, and visualization to make sense of these huge 3D datasets (Aim 3). Synchrotron-based x-ray microtomography has been unknown to most neuroscientists. It fills the information and resolution gap between MRI studies of living animals and humans, light microscopy of specific molecule types within a few millimeters of the brain surface, and electron microscopy studies with exquisite anatomical detail of limited regions. Our team includes experts in brain tissue preparation and electron microscopy, x-ray microtomography, and analysis of brain anatomy. We are therefore in a unique position to develop x-ray tomography for massive scale brain anatomy, and to make these advances available to the neuroscience community since the APS is a no-cost user facility.

Abstract Depression is common, costly, and a leading cause of disability. Assessment of behavior and experience related to depression has tended to rely on self-report and interview-based methods. Environmental momentary assessment inserts assessment into people's lives, but still requires active engagement by those being evaluated. We propose to develop and validate a mobile phone-based personal sensing system to detect depression and related behaviors that relies on sensor data that are collected continuously and unobtrusively. Because people tend to keep their phones with them, the mobile phone is an ideal sensing platform, as it can continuously collect data in the context of the individual's life with no ongoing effort on the part of the user. Such systems are already being used to detect simple behaviors, such as activity recognition and sleep quantification, which are more proximal to the sensor data. Aim 1 will develop markers for a broad range of behavioral targets related to symptoms of major depressive episode (MDE; anhedonia, negative mood, sleep disruption, psychomotor activity, fatigue, and diminished concentration) and related domains (e.g. social functioning, stress, motivation) across a representative sample of participants. Aim 2 will combine all behavioral targets using machine learning to 1) estimate MDE and symptom severity cross-sectionally, 2) identify transition from non-depressed to depressed states, and depressed to non-depressed states, and 3) predict MDE and symptom severity 4 and 8 weeks out. Aim 3 will seek to understand the complex relationships among behavioral targets and depression. We will accomplish this by enrolling 1200 representative participants, in six 4-month waves of data collection. Each participant will download software that collects a wide variety of sensor data (GPS, accelerometry, light, Bluetooth, phone usage, etc.) and an app that collects ecological momentary assessments (EMA). Following each wave we will develop algorithms for a subset of behavioral targets and features (a definition of raw sensor data that incorporates meaning, like translating GPS data into ?home?). Each algorithm will then be validated in the subsequent wave. After 5 waves (1000 participants), the set of all markers of behavioral targets and features will be combined using machine learning to detect and predict depression. This hierarchical approach extracts information from data at multiple levels, which ultimately is far more likely to succeed than relying solely on raw sensor data. The final wave will serve to replicate and validate the entire depression prediction model. This sensing platform is scientifically significant, as it will provide a fundamentally new tool for obtaining continuous, objective markers of behavior that are relevant to depression, as well as many other psychiatric and medical disorders. This project has the potential develop new understandings into the etiology of depression. It is clinically significant, as it will allow for continuous, effortless assessment of populations at risk for depression and ongoing evaluation during treatment.

NEURONAL CIRCUITS FOR CONTEXT-DRIVEN BIAS IN AUDITORY CATEGORIZATION In everyday life, because both sensory signals and neuronal responses are noisy, important cognitive tasks, such as auditory categorization, are based on uncertain information. To overcome this limitation, listeners incorporate other types of signals, such as the statistics of sounds over short and long time scales and signals from other sensory modalities into their categorization decision processes. At the behavioral level, such contextual signals bias categorization by shifting the listener's psychometric curve. At the neuronal level, categorization requires a transformation of sensory representation into a representation of category membership that is modulated by these contextual signals. While categorical representations have been found in the cortex, the cell types and neuronal mechanisms supporting the emergence of these representations remains unknown. Furthermore, the mechanisms by which neuronal categorical representations are modulated by contextual signals, giving rise to a behavioral bias, have not been explored. Our goal is to identify the contribution of specific cell types to categorization and to understand the neuronal mechanisms for how contextual signals bias auditory categorization. Multiple studies have demonstrated that neurons in auditory cortex (AC) and the posterior parietal cortex (PPC) are involved in auditory categorization. Based on the well-described circuit architecture of the AC, recent studies, and our preliminary data, we propose a series of hypotheses that delineate the role of excitatory-inhibitory circuits within AC in creating and biasing categorical stimulus representations and for the role of PPC-AC projections in driving the source for the bias signal. To test these hypotheses, we train mice in a two-alternative-forced choice task in which mice categorize the task, associations). frequency of a ?target? sound into one of two overlapping categories (?low? or ?high?). While mice participate in this we systematically manipulate three bias signals (short-term and long-term stimulus statistics, and cross-modal Thisdesign allows us to frame the cognitive task within a Bayesian framework, which generates formal computational models for the function of specific neuronal cell types that are tested experimentally. behavioral activity. category. in auditory We will combine this and computational framework with electrophysiological recordings and optogenetic manipulations of neuronal First, we will test whether distinct neuronal cell types in AC differentially encode information about stimulus Second, we will test whether and how specific inhibitory neuronal cell types in AC mediate context dependence auditory categorization. Third, we will test whether and how cortico-cortical feedback mediates context dependence in categorization. Aligned with the goals of the BRAIN initiative, our project will deliver a mechanistic framework for a cortical circuit supporting a complex behavior. These results will quantitatively address an important open question to what extent the same or distinct neuronal populations integrate information across multiple temporal scales and across sensory modalities, generalizing or specializing the representation of the bias in categorization.

Artificial intelligence need not replicate the human brain, but may constructively take inspiration from it. The brain is organized hierarchically, from large brain structures to smaller regions, cortical columns, all the way down to microcircuits made of a few interacting neurons. These modules may make our brains more efficient by dissecting problems into locally solvable subproblems. This in turn may make us learn faster by figuring out where in the brain a mistake was made and allow us to do better on new problems. Modules may also make the brain more efficient, allowing it to use fewer neurons and synapses. In those areas where brains seem to benefit from modularity, modern deep learning systems appear to be weaker. Building modules into deep neural networks promises to greatly improve generalization, interpretability, credit assignment in learning, computational cost and make them more resilient to adversarial stimuli. The result of this project will be improvements to the performance and understanding of modern artificial intelligence systems. The project will contribute in a broad way to the dissemination of computational results to neuroscience and of neuroscience results to the computational community through a combination of summer schools, teaching, and publishing.

To meet these goals, this project enables a broadly interdisciplinary approach both to produce systems with modularity and to dissect their modular aspects. The research aims to build networks that, while learning, dissect training tasks by incrementally developing structural modules. This is done by minimizing cost functions that evaluate the community structure of neuron connectivity. It will design learning algorithms that encourage modularity by performing credit assignment at the level of modules. This is done by gating learning at both the neuron and the module level. Finally, the project team will develop tools for interpreting modular networks. This is done by performing psychophysical experiments with human subjects and by computational analysis. All of these components interrelate, provide a unified picture of how modularity can improve current machine learning approaches, and build a new bridge between neuroscience and deep learning.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

Abstract: Causality is central to neuroscience. For example, we might ask about the causal effect of a neuron on another neuron, or its influence on perception, action, or cognition. Moreover, any medical approaches aim at producing a causal effect ? effecting improvements for patients. Randomized controlled trials (RCTs) are the gold standard to establish causality, but they are not always practical. For example, while we can electrically or optogenetically activate entire areas, large-scale targeted stimulation of individual neurons is hard. Other ways of establishing causality are problematic: if we observe a correlation it is hard to know its cause. The problem is confounding: there are variables that we do not record that affect the variables we do. This also renders model comparisons problematic ? a causally wrong model with few parameters may well fit the observed data better than a causally correct one with many parameters. We thus need data analysis tools that allow authoritatively asking causal questions without the need for random perturbation experiments. Just like neuroscience now, the field of econometrics once focused on correlations. But since the 1980s, empirical economics has undergone a so-called credibility revolution, requiring the development of rigorous methods to establish causality. Several successful methods have emerged to become the workhorses of empirical economics. The idea underlying these methods is that if one can observe variables that approximate random perturbations, then one can still discover causal relations. This is what economists call a quasi-experiment. We here propose to carry over such quasi-experimental techniques to neuroscience. For example in neuroscience, if there is a random variable that affects only one neuron, then any activity in other neurons correlated with that variable must be causally affected by the neuron. Another famous quasi- experimental method is regression discontinuity design (RDD). This approach effectively uses the noise introduced at the threshold to identify causal relations. Importantly, such techniques have, thanks to decades of research in econometrics, very well understood statistical properties. These approaches promise to considerably enrich the approaches towards causality we have in neuroscience. We have a strong interdisciplinary team, spanning economics, experimental, and computational neuroscience, collaborating on adapting these quasi-experimental techniques to problems in neuroscience through a combination of machine learning and domain-specific engineering. This promises to be a major advance relative to current techniques that generally approach causality in neuroscience through model comparison.

PROJECT SUMMARY/ABSTRACT The overall goal of this R01 project is to develop an automated assessment system that can capitalize on state of the art sensing technologies and machine learning algorithms to enable accurate and early detection of infants at risk for neurodevelopmental disabilities. In the USA, 1 in 10 infants are born at risk for these disabilities. For children with neurodevelopmental disabilities, early treatment in the first year of life improves long-term outcomes. However, we are currently held back by inadequacies of available clinical tests to measure and predict impairment. Existing tests are hard to administer, require specialized training, and have limited long- term predictive value. There is a critical need to develop an objective, accurate, easy-to-use tool for the early prediction of long-term physical disability. The field of pediatrics and infant development would greatly benefit from a quantitative score that would correlate with existing clinical measures used today to detect movement impairments in very young infants. To realize a new generation of tests that will be easy to administer, we will obtain large datasets of infants playing in an instrumented gym or simply being recorded while moving in a supine posture. Video and sensor data analyses will convert movement into feature vectors based on our knowledge of the problem domain. Our approach will use machine learning to relate these feature vectors to currently recommended clinical tests or other ground truth information. The power of this design is that algorithms can utilize many aspects of movement to produce the relevant scores. Our preliminary data allows us to lay the following aims: 1)Aim 1: To assess concurrent validity of a multimodal instrumented gym with existing clinical tools. Here, using 150 infants (75 with early brain injury and 75 controls), we will focus on converting data from an instrumented gym into estimates of the standard clinical tests; 2)Aim 2: To develop a computer vision-based algorithm to quantify infant motor performance from single camera video. Here using video data from 1200 infants (400 with early brain injury, 400 preterm without early brain injury, 400 controls), plus those gathered from Aim 1 and Aim 3, we will extract pose data from single-camera video recordings and convert these into kinematic features and relevant scores needed to classify infant movement; 3)Aim3: To discover the features related to long-term motor development. Here we will convert data collected longitudinally from 50 infants (25 with early brain injury and 25 controls) using both instrumented gym and video recordings into estimates standard clinical tests change over time and track features over developmental timescales. These three aims spearhead the use of real world behavior for movement scoring. Our aims will bring us closer to a universal non-invasive test for early detection of neurodevelopmental disabilities and lay the groundwork for long-term prediction of disability. But above all, it promises to scale to infants worldwide, producing an affordable tool to aid in infant health assessment.

Artificial Intelligence (AI) is driving considerable parts of the US economy, influencing almost every industrial and scientific sector. Yet, the "intelligence" of these systems still cannot match the flexibility and breadth of even simple biological systems. The objective of this project is to develop a revolutionary new class of AI by focusing on four insights from the biological intelligence (BI) of animals that, unlike current AI agents, (1) do not start as blank slates (2) do not forget when they learn new things (3) have curiosity, and (4) interpret the world in terms of cause and effect. This project brings together outstanding scientists from a wide variety of disciplines and diverse backgrounds to tackle this problem. Through planning meetings and collaborative exercises, the team will generate preliminary data, and preparatory work for a future center focused on reframing the fundamental question of "intelligence." Outreach efforts will engage many diverse participants through large-scale online teaching.

At the founding of AI, Alan Turing proposed a test to determine when an AI behaves like a BI, and specifically human intelligence (HI). This test set the stage for the following 70 years of AI development. It has largely been replaced by narrow competitions aimed at imitating humans at specific tasks, such as playing certain games, identifying objects, or translating languages. But neither the Turing test nor today?s AI competitions utilize modern conceptualizations of what it means to be intelligent. It has become clear that intelligence evolved to incorporate several complex capabilities, which are critically missing from today?s AI: (1) preprogramming biases and baseline behaviors, (2) continually leveraging of many previous experiences to improve decision making in newly encountered tasks, (3) actively seeking out information that is useful for future decisions even if these differ from past decisions, and, finally, (4) constructing causal models relevant to decisions and communicating these models. This project brings together a unique group that truly understands both AI and BI/HI to address this gap. This group aims to define what is missing in AI relative to BI/HI, and to determine which research paths can enhance future approaches. In year one, the project will develop a test to measure a specific aspect of intelligence found in animals, but not current AI. This test will be sufficiently simple that it can be performed by several different biological taxa, humans, as well as AI. In year two, the participants will conduct pilot experiments to quantify current levels of performance on these tests and distill insights from BI/HI for AI. The resulting benchmarks will provide explicit and quantitative milestones for the eventual institute, whose goal will be to develop AI that matches HI on the new tests of intelligence.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

Neuromatch Academy (NMA) is an online summer school that will train about 7,000 neuroscientists in modern data science and modeling methods. These methods are the basis of considerable endeavors in industry and academia, including neuroscience, robotics, and cognitive modeling. This project will support student participation and evaluation of the academy.

A central component of the education of neuroscientists is intensive summer schools, which usually run for two weeks. They are usually expensive, require travel, and open only to small numbers of participants. NMA aims to provide a three-week summer school in an online fashion providing outstanding quality of education. It uses teaching-assistant-led small virtual tutorials (about 10 students each), globally broadcast question and answer sessions through crowdcast, professor-guided project work, and a host of further approaches to produce a deep learning experience. Focusing on a tight interactions with other participants, TAs, and professors give it a unique profile among approaches to online teaching. In a post-COVID world, innovating in online education is of crucial importance and NMA has both a strong mission of innovation and a strong evaluation plan to make sure it delivers on its promise. The evaluation approach will consists of questionnaires sent to participants, TAs, and lecturers as well as a causal analysis to see how NMA affects subsequent text use and collaboration patterns.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.