Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2019 / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 179-192). / From riding a bike to brushing our teeth, we learn many of our motor skills through trial and error. Many biologically based trial and error learning models depend on a teaching signal from dopamine neurons. Dopamine neurons increase their firing rates to signal outcomes that are better than expected and decrease their firing rates to signal outcomes that are worse than expected. This dopamine signal is thought to control learning by triggering synaptic changes in the basal ganglia. What are the origins of this dopaminergic teaching signal? How do synaptic changes in the basal ganglia lead to changes in behavior? In this thesis, I study these questions in a model of skill learning - the songbird. In the first part of my thesis, I develop a computational model of song learning. This model incorporates a dopaminergic reinforcement signal in VTA and dopamine-dependent synaptic plasticity in the singing-related part of the basal ganglia. / I demonstrate that this model can provide explanations for a variety of experimental results from the literature. In the second part of my thesis, I investigate a potential source of the dopaminergic error signal in VTA. I performed the first recordings from one cortical input to VTA: the dorsal intermediate arcopallium (AId). Previous studies disagree on the role of Ald in behavior. Some studies argue that AId contributes vocal error information to VTA. Other studies suggest that AId is not involved in the computation of error signals, but is instead responsible for controlling head and body movements. I directly tested these hypotheses by recording single neurons in AId during singing and during natural movements. My results support a motor role for AId - AId neurons had highly significant changes in activity during head and body movements. Meanwhile, following vocal errors Aid neurons had small but marginally significant decrease in firing rate. / In a more detailed analysis, I developed an automated behavior classification algorithm to categorize zebra finch behavior and related these behavior classes to the activity of single units in Aid. My results support the hypothesis that AId is part of a general-purpose motor control network in the songbird brain. / by Michael E. Stetner. / Ph. D. / Ph.D. Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2019 / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 123-130). / Humans are remarkable in their ability to rapidly learn complex tasks from little experience. Recent successes in Al have produced algorithms that can perform complex tasks well in environments whose simple dynamics are known in advance, as well as models that can learn to perform expertly in unknown environments after a great amount of experience. Despite this, no current AI models are able to learn sufficiently rich and general representations so as to support rapid, human-level learning on new, complex, tasks. This thesis examines some of the epistemic practices, representations, and algorithms that we believe underlie humans' ability to quickly learn about their world and to deploy that understanding to achieve their aims. In particular, the thesis examines humans' ability to effectively query their environment for information that helps distinguish between competing hypotheses (Chapter 2); children's ability to use higher-level amodal features of data to match causes and effects (Chapter 3); and adult human rapid-learning abilities in artificial video-game environments (Chapter 4). The thesis culminates by presenting and testing a model, inspired by human inductive biases and epistemic practices, that learns to perform complex video-game tasks at human levels with human-level amounts of experience (Chapter 5). The model is an instantiation of a more general approach, Theory-Based Reinforcement Learning, which we believe can underlie the development of human-level agents that may eventually learn and act adaptively in the real world. / by Pedro A. Tsividis. / Ph. D. / Ph.D. Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2019 / Cataloged from PDF version of thesis. "Some pages in the original document contain text that runs off the edge of the page. See Appendix A - pages 162-171"--Disclaimer Notice page. / Includes bibliographical references (pages 147-159). / It is commonly assumed that there is a reliable one-to-one mapping between a certain configuration of facial movements and the specific emotional state that is supposedly signals. One common way to test this one-to-one hypothesis is to ask people to deliberately pose the facial configurations that they believe they use to express emotions. Participants are randomly sampled, without concern for their emotional expertise, and are given a single emotion word or a single, brief statement to describe each emotion category. They then deliberately pose the facial configuration that they believe they make when expressing instances of this category. Such studies routinely find that participants from different countries show moderate to strong evidence for a one-to-one mapping between an emotion category and a single facial configuration (its presumed facial expression). / In Study 1, we examined the facial configurations posed by emotion experts - famous actors who were provided with a diverse sample of richly described scenarios, full of context. Participants inferred the emotional meaning of the scenarios, which were then grouped into categories. Systematic coding of the facial poses for each emotion category revealed little evidence for the hypothesis that each category has a diagnostic facial expression. Instead, we observed a high degree of variability among expert's facial poses for any given emotion category, and little specificity for any pose. Furthermore, an unsupervised statistical analysis discovered 29 novel emotion categories with moderately consistent facial poses. In Study 2, participants were asked to infer the emotional meaning of each facial pose when presented alone, or when presented in the context of its eliciting scenario. / In fact, the majority of studies designed to test the one-to-one hypothesis ask people from various cultures to judge posed configurations of facial movements, such as a scowl (the proposed facial expression for anger), a frown (the proposed expression for sadness), and so on, on the assumption that these facial configurations, as universal expressions of emotional states, co-evolved with the ability to recognize and read them. These studies routinely show participants one facial configuration posed by multiple posers for each emotion category and observe variable findings, depending on the experimental method used. Our analyses indicated that participants's inferences about the emotional meaning of the facial poses were influenced more by their eliciting scenarios than by the physical morphology of the facial configurations. / These findings strongly replicate emerging evidence that the emotional meaning of any set of facial movements may be much more variable and context-dependent than hypothesized by the common one-to-one view which continues to influence the public understanding of emotion, and hence education, clinical practice, and applications in government and industry. Although more ecologically valid research on how people actually move their faces to express emotion is urgently needed, doing so was immensely difficult without the right tools that support the process of capturing facial data in real life, automatically processing these data, and finally supporting data verification and analysis. We developed a system of technological tools to support the investigations of facial movements during emotional episodes in naturalistic settings with the use of dynamic and longitudinal facial data. We then collected, pre-processed, verified and analyzed data from Youtube using our newly-developed tools. / In particular, we examined two talk show hosts and presented preliminary insights on the answers to questions that were previously very difficult to investigate. / by Tuan Le Mau. / Ph. D. / Ph.D. Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2019 / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 91-100). / Neural networks and probabilistic models have different and in many ways complementary strengths and weaknesses: neural networks are flexible and support efficient inference, but rely on large quantities of labeled training data. Probabilistic models can learn from fewer examples, but in many cases remain limited by time-consuming inference algorithms. Thus, both classes of models have drawbacks that both limit their engineering applications and prevent them from being fully satisfying as process models of human learning. This thesis aims to address this state of affairs from both directions, exploring case studies where we make neural networks that learn from less data, and in which we design more efficient inference procedures for generative models. First, we explore recurrent neural networks that learn list-processing procedures (sort, reverse, etc.), and show how ideas from type theory and programming language theory can be used to design a data augmentation scheme that enables effective learning from small datasets. Next, we show how error-driven proposal mechanisms can speed up stochastic search for generative model inversion, first developing a symbolic model for inferring Boolean functions and Horn clause theories, and then a general-purpose neural network model for doing inference in continuous domains such as inverse graphics. / by Owen Lewis. / Ph. D. / Ph.D. Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2014. / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 211-220). / People can learn a new concept almost perfectly from just a single example, yet machine learning algorithms typically require hundreds or thousands of examples to perform similarly. People can also use their learned concepts in richer ways than conventional machine learning systems - for action, imagination, and explanation suggesting that concepts are far more than a set of features, exemplars, or rules, the most popular forms of representation in machine learning and traditional models of concept learning. For those interested in better understanding this human ability, or in closing the gap between humans and machines, the key computational questions are the same: How do people learn new concepts from just one or a few examples? And how do people learn such abstract, rich, and flexible representations? An even greater puzzle arises by putting these two questions together: How do people learn such rich concepts from just one or a few examples? This thesis investigates concept learning as a form of Bayesian program induction, where learning involves selecting a structured procedure that best generates the examples from a category. I introduce a computational framework that utilizes the principles of compositionality, causality, and learning-to-learn to learn good programs from just one or a handful of examples of a new concept. New conceptual representations can be learned compositionally from pieces of related concepts, where the pieces reflect real part structure in the underlying causal process that generates category examples. This approach is evaluated on a number of natural concept learning tasks where humans and machines can be compared side-by-side. Chapter 2 introduces a large-scale data set of novel, simple visual concepts for studying concept learning from sparse data. People were asked to produce new examples of over 1600 novel categories, revealing consistent structure in the generative programs that people used. Initial experiments also show that this structure is useful for one-shot classification. Chapter 3 introduces the computational framework called Hierarchical Bayesian Program Learning, and Chapters 4 and 5 compare humans and machines on six tasks that cover a range of natural conceptual abilities. On a challenging one-shot classification task, the computational model achieves human-level performance while also outperforming several recent deep learning models. Visual "Turing test" experiments were used to compare humans and machines on more creative conceptual abilities, including generating new category examples, predicting latent causal structure, generating new concepts from related concepts, and freely generating new concepts. In each case, fewer than twenty-five percent of judges could reliably distinguish the human behavior from the machine behavior, showing that the model can generalize in ways similar to human performance. A range of comparisons with lesioned models and alternative modeling frameworks reveal that three key ingredients - compositionality, causality, and learning-to-learn - contribute to performance in each of the six tasks. This conclusion is further supported by the results of Chapter 6, where a computational model using only two of these three principles was evaluated on the one-shot learning of new spoken words. Learning programs with these ingredients is a promising route towards more humanlike concept learning in machines. / by Brenden M. Lake. / Ph. D.

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2014. / Cataloged from PDF version of thesis. / Includes bibliographical references. / Neuroscientists hold widely divergent opinions on the behavioral relevance of oscillatory brain states. Some consider them to be a side effect of anatomical connectivity, with little or no role in guiding action. Others view them as a fundamental feature of the network states that underlie perception and cognition. In this thesis, I take a systematic approach to studying two of the most prominent types of oscillations,'gamma rhythms in the neocortex (30-80 Hz) and theta rhythms in the hippocampus (4-12 Hz). In both cases, I use light-gated ion channels to manipulate spike activity on a cycle-by-cycle basis in awake, behaving mice. By rhythmically stimulating fast-spiking interneurons in somatosensory cortex, I can emulate the activity patterns that define gamma oscillations under natural conditions. Emulating gamma enhances the detection of threshold-level vibrissae deflections, analogous to the behavioral effects of shifting attention. By triggering stimulation of fast-spiking interneurons in the hippocampus on peaks and troughs of endogenous rhythms, I can reduce spike activity at specific phases of theta. In the context of a spatial navigation task, I find that the ability of inhibition to enhance decision-making accuracy depends on both the theta phase and the task segment in which it occurs. Both of these experiments provide novel causal evidence for the behavioral impact of oscillations, which offers a much more compelling argument for their utility than traditional correlative measures. Finally, I present a new platform for extracellular electrophysiology. This platform, called Open Ephys, makes the closed-loop experiments that are ideal for studying oscillations accessible to a wider audience. / by Joshua H. Siegle. / Ph. D.

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2014. / Cataloged from PDF version of thesis. / Includes bibliographical references. / During sleep and general anesthesia, the brain enters a state of decreased arousal and consciousness is transiently suspended. How this transition occurs is a fundamental and unsolved question in neuroscience. The neural dynamics that disrupt consciousness have not been identified, and the circuit mechanisms that generate these dynamics remain unknown. Furthermore, understanding the neural basis of sleep and anesthesia is key to improving clinical monitoring of patients undergoing general anesthesia and to advancing treatments of sleep disorders and neurological conditions such as coma. In this thesis, I combine intracranial electrophysiology in human subjects with optogenetic manipulation of thalamocortical circuits in mice to identify the neural dynamics underlying sleep and anesthesia. I first show that loss of consciousness during propofol general anesthesia is associated with the abrupt onset of slow oscillations that disrupt cortical networks. I then demonstrate that activation of the thalamic reticular nucleus generates slow wave activity and decreases arousal state, identifying a causal mechanism that generates physiological and behavioral signs of sleep. Finally, I study patients undergoing deep general anesthesia at levels corresponding to medically induced coma, and show that this state is marked by local cortical dynamics consistent with impaired cerebral metabolism. Taken together, these results identify a set of neural dynamics associated with unconscious states, and demonstrate specific mechanisms for how they disrupt brain function. These findings provide new insight into the neuroscience of arousal states, and suggest clinical approaches that could improve patient care. / by Laura D. Lewis. / Ph. D.

Thesis (Ph.D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 1998. / Includes bibliographical references (p. 115-122). / An experiment was conducted to determine whether tests used to assess working memory in different disciplines (neuroimaging, psycholinguistics, neuropsychology) are highly correlated, and thus whether they are equivalent measures of a unitary underlying function. Scores on the different tests (N-back, reading span, backward digit span) did not correlate highly, and were predicted by measures of different hypothesized components of working memory. These results indicate that working memory is best conceived of as a system of multiple, interacting components that contribute to different aspects of task performance, rather than as a single, unified resource, and that currently popular tests of working memory cannot be used interchangeably to measure working memory. A second experiment was conducted to examine the relation between sentence memory and working memory, and to determine whether memory for sentences is a function of the number of clauses in the sentence, or the number of new discourse referents. Subjects heard sentences of different lengths (2 - 5 clauses) and structures (relative clause, sentential complement, double object). Double object sentences contained one additional discourse referent per clause than the other two sentence types. / (cont.) If new discourse referents are the units of sentence memory, performance should be worse on double object sentences. If clauses are the unit of sentence memory, accuracy should be the same for all three sentence types. There were no reliable differences between double object sentences and the other two sentences types, indicating the clauses are the units of sentence memory. Subjects recalled 2-clause sentences highly accurately, and recalled 4-clause and 5-clause sentences poorly. There were large individual differences in the recall of 3-clause sentences. Over half of this variance was accounted for by individual differences in working memory. Measures of two hypothesized working memory components, the central executive and the short-term store, each accounted for independent variance in the sentence memory score. / by Rose M. Roberts. / Ph.D.

Thesis: S.M. in Neuroscience, Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2018. / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 27-29). / Reconstruction of the 3D morphology of neurons is an essential step towards the analysis and understanding of the structures and functions of neural circuits. Optical microscopy has been a key technology for mapping brain circuits throughout the history of neuroscience due to its low cost, wide usage in biological sciences and ability to read out information-rich molecular signals. However, conventional optical microscopes have limited spatial resolution due to the diffraction limit of light, requiring a tradeoff between the density of neurons to be imaged and the ability to resolve finer structures. A new technology called expansion microscopy (ExM), which physically expands the specimen multiple times before optical imaging, enables us to image fine structures of biological tissues with super resolution using conventional optical microscopes. With the help of ExM, we can also read out molecular information in the neural tissues optically. In this thesis, I will introduce our study of the properties, via computer simulation, of a candidate automated approach to algorithmic reconstruction of dense neural morphology, based on simulated data of the kind that would be obtained via two emerging molecular technologies-expansion microscopy (ExM) and in-situ molecular barcoding. We utilize a convolutional neural network to detect neuronal boundaries from protein-tagged plasma membrane images obtained via ExM, as well as a subsequent supervoxel-merging pipeline guided by optical readout of information-rich, cell-specific nucleic acid barcodes. We attempt to use conservative imaging and labeling parameters, with the goal of establishing a baseline case that points to the potential feasibility of optical circuit reconstruction, leaving open the possibility of higher-performance labeling technologies and algorithms. We find that, even with these conservative assumptions, an all-optical approach to dense neural morphology reconstruction may be possible via the proposed algorithmic framework. / by Peilun Dai. / S.M. in Neuroscience

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 1992. / Includes bibliographical references (leaves 116-130). / by Kyoungmin Lee. / Ph.D.