Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 2009. / This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections. / Includes bibliographical references (p. 97-103). / It is the goal of this thesis to examine the frontal cortex-basal ganglia system during arbitrary visuomotor association learning, the forming of arbitrary links between visual stimuli and motor responses (e.g. red means stop), a fundamental learning process that underlies much of our complex behavior such as written language. The experiments contained in this thesis investigate the involvement of four components of this system in the acquisition of these associations: dorsolateral prefrontal cortex (dlPFC), caudate nucleus (Cd), frontal eye field (FEF), and the internal segment of the globus pallidus (GPi). Extracellular electrophysiological recordings were performed in awake-behaving primates performing three different learning tasks. In the different behavioral paradigms used in these studies, learning with and without reversals is investigated and compared both directly within the same experiment and indirectly across experiments. The results of these studies suggest that a complex interplay between brain areas in the frontal cortex-basal ganglia system exists. The study of FEF during reversal learning revealed that FEF contains task-related information from the start of learning, suggesting that it may be passing information onto PFC and Cd to aid the learning process. In addition, GPi is shown to contain more specific information about the learned association during the reversal task, providing evidence for an increase in the complexity of information processing through the basal ganglia. / (cont.) The in-depth study of dlPFC and Cd suggests that the frontal cortex-basal ganglia system functions only when competition between learning contexts exist. When all competition is eliminated by removing reversal learning from the behavioral task, Cd does not show involvement in the learning process. But when competition exists, both Cd and PFC show learning-related changes in task-relevant information. As determined by coherence analysis of local field potentials, communication between dlPFC and Cd is greater during reversal learning, when competition is heightened. This communication also decreases as learning progresses suggesting a role in the transfer of information between areas in facilitating the learning process. Overall, these studies further the understanding of the role of the frontal cortex-basal ganglia system in arbitrary visuomotor learning and posit that the function of the system is dependent on the existence of competition between learned information. / by Michelle S. Machon. / Ph.D.

Thesis (S.M.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 2003. / Includes bibliographical references (v. 2, leaves 268-271). / (cont.) the functionality of the fMRI data analysis and visualization tools used at Massachusetts General Hospital. I analyze and interpret an 18-subject fMRI experiment I ran using the new task design and software tools. Finally, I present preliminary findings on linguistic questions as well as the nature of fMRI signal, using direct Electrophysiological data recorded from electrodes implanted in the brains of two Epilepsy patients. These patients had electrodes implanted through or near classical language areas of their brains, as a necessary clinical step in locating and surgically removing the seizure-causing tissue. The main findings of this thesis are: 1.) Morphology alone can activate Broca's area, 2.) Other areas are involved, including BA47, anterior insula, and SMA, 3.) Broca's area and BA47 respond to application of abstract grammatical features, even without phonological manipulations, 4.) Morphophonological manipulation additionally recruits insula and SMA, 5.) While simply accessing nouns versus verbs may involve separable brain regions, inflectional processing of the two categories may be done by the same process, 6.) Regularly and Irregularly inflected verbs show a double dissociation of activation in frontal and medial regions, 7.) Processing of English noun more than verb morphology may rely on some contribution from number processing brain systems ... / Inflectional morphology is the component of language concerned with changing a word's form to reflect context-specific meaning, such as the affixing of"-ed" for English verbs in the past tense, or adding "-s" in order to signal a noun's plural form. Although it is but one part of language, morphology may be useful as a model system for larger issues in language and cognition. Morphological processing touches on: the manipulation of memorized items (the vocabulary of words, and maybe word endings), the application and power of combinatorial rules (to generate correct forms, even of unknown words), and the binding of units of information into meaning. Morphology's relationship with other more traditional facets of language such as syntax (sentence structure) and semantics (meanings of individual words) is debated, as is the objective reality of grammatical categories (e.g. noun/verb) as well as combinatorial rules. Functional Magnetic Resonance Imaging (fMRI) is an exciting technique for peering into the brain and answering questions about its function. However, the technique has limited temporal and spatial resolution, and indexes the brain basis of cognition only indirectly, via blood response to cellular metabolism. In this thesis I propose a task for manipulating morphological production, embedded in a 2*2*3 design simultaneously varying grammatical class (nouns versus verbs), regularity of inflection (e.g. words like "walk" which take the stereotyped or regular ending "-ed", versus those like "bring" and "sing" which have idiosyncratic past tense forms), and three types of morphological task aimed to separate the assignment of grammatical features (e.g. present/past tense) from changes in word sound. I introduce and utilize software to extend / by Nedim T. Sahin. / S.M.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 2009. / "June 2009." Cataloged from PDF version of thesis. / Includes bibliographical references. / Unlike the human brain, which produces few neurons in adulthood, the brains of songbirds continue to produce new neurons throughout life. The function of these new neurons is not know, although it has been suggested that they endow the avian brain with a remarkable regenerative capacity that does not exist in mammals. It has also been proposed that the addition of new neurons in adulthood underlies behavioral plasticity, such as song learning. A better understanding of the cellular mechanisms that control the addition of new neurons to the postnatal brain may help clarify its biological function. This thesis is an investigation of the cell biology of postnatal neurogenesis in the songbird forebrain, with special emphasis on the High Vocal Center. Neuronal progenitors in the juvenile zebra finch brain were identified by fate mapping using engineered retroviruses. Multiple populations of neural progenitors appear to exist in the juvenile zebra finch brain, and each produces different types of neurons. At least three cell types appear to be added to the postnatal finch brain. Homology between neurogenesis in the postnatal finch and embryonic mammalian forebrain was also assessed. To characterize the mechanism of cell addition, videos were made, documenting the migration and integration of new neurons into the High Vocal Center. Neural progenitors were labeled using retroviruses, carrying the gene for the green fluorescent protein, allowing new neurons to be observed in the intact brain, with a powerful infrared laser. By replacing a small hole in the skull with a piece of optical glass, one could observe labeled neurons periodically over many days as they were born until they wired up to the existing circuitry. New neurons engaged in a previously undescribed form of migration. Further study of this form of neuron migration as well as other aspects of postnatal neurogenesis may lead to the development of strategies for replacing neurons in the human brain lost to death or disease. / by Benjamin Barnett Scott. / Ph.D.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 2012. / This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections. / Cataloged from student-submitted PDF version of thesis. Page 125 blank. / Includes bibliographical references. / The integration of anatomical, functional, and developmental approaches in cognitive neuroscience is essential for generating mechanistic explanations of brain function. In this thesis, I first establish a proof-of-principle that neuroanatomical connectivity, as measured with diffusion weighted imaging (DWI), can be used to calculate connectional fingerprints that are sufficient to delineate fine anatomical distinctions in the human brain (Chapter 2). Next, I describe the maturation of structural connectivity patterns by applying these connectional fingerprints to over a hundred participants ranging from five to thirty years of age, and show that these connectional patterns have different developmental trajectories (Chapter 3). I then illustrate how anatomical connections may shape (or in turn be shaped by) function and behavior, within the framework of reading ability and describe how white matter tract integrity may predict future acquisition of reading ability in children (Chapter 4). I conclude by summarizing how these experiments offer testable hypotheses of the maturation of structure and function. Studying the complex interplay between structure, function, and development will get us closer to understanding both the constraints present at birth, and the effect of experience, on the biological mechanisms underlying brain function. / by Zeynep Mevhibe Saygin. / Ph.D.

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2016. / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 117-129). / I argue that the intersection of deep learning, hierarchical reinforcement learning, and generative models provides a promising avenue towards building agents that learn to produce goal-directed behavior given sensations. I present models and algorithms that learn from raw observations and will emphasize on minimizing their sample complexity and number of training steps required for convergence. To this end, I introduce hierarchical variants of deep reinforcement learning algorithms, which produce and utilize temporally extended abstractions over actions. I also present a hybrid model-free and model-based deep reinforcement learning model, which can also be potentially used to automatically extract subgoals for bootstrapping temporal abstractions. I will then present a model-based approach for perception, which unifies deep learning and probabilistic models, to learn powerful representations of images without labeled data or external rewards. Learning goal-directed behavior with sparse and delayed rewards is a fundamental challenge for reinforcement learning algorithms. The primary difficulty arises due to insufficient exploration, resulting in an agent being unable to learn robust value functions. I present the Deep Hierarchical Reinforcement Learning (h-DQN) approach, which integrates hierarchical value functions operating at different time scales, along with goal-driven intrinsically motivated behavior for efficient exploration. Intrinsically motivated agents can explore new behavior for its own sake rather than to directly solve problems. Such intrinsic behaviors could eventually help the agent solve tasks posed by the environment. h-DQN allows for flexible goal specifications, such as functions over entities and relations. This provides an efficient space for exploration in complicated environments. I will demonstrate h-DQN's ability to learn optimal behavior given raw pixels in environments with very sparse and delayed feedback. I will then introduce the Deep Successor Reinforcement (DSR) learning approach. DSR is a hybrid model-free and model-based RL algorithm. It learns the value function of a state by taking the inner product between the state's expected future feature occupancy and the corresponding immediate rewards. This factorization of the value function has several appealing properties - increased sensitivity to changes in the reward structure and potentially the ability to automatically extract subgoals for learning temporal abstractions. Finally, I argue for the need for better representations of images, both in reinforcement learning tasks and in general. Existing deep learning approaches learn useful representations given lots of labeled data or rewards. Moreover, they also lack the inductive biases needed to disentangle causal structure in images such as objects, shape, pose and other intrinsic scene properties. I present generative models of vision, often referred to as analysis-by-synthesis approaches, by combining deep generative methods with probabilistic modeling. This approach aims to learn structured representations of images given raw observations. I argue that such intermediate representations will be crucial to scale-up deep reinforcement learning algorithms, and to bridge the gap between machine and human learning. / by Tejas Dattatraya Kulkarni. / Ph. D.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 2010. / Cataloged from PDF version of thesis. / Includes bibliographical references. / The basal ganglia, and specifically the sensorimotor (dorsolateral) striatum, have been implicated in stimulus-response learning. Here, I analyze the role the striatum plays in learning. We recorded from neurons of the sensorimotor striatum as rats learn, are over-trained, are extinguished, and are re-trained on a discriminative T-maze task. In this T-maze a gate was lowered immediately after an auditory click and the rats were allowed to proceed down the long arm of the maze. Mid-run, one of two tones was played. Rats had to choose to turn down either the left or right arm of the T-maze based on which tone was played. We discovered that population neural activity becomes restructured during learning and overtraining to emphasize the beginning and end of each trial. We also created a short-term memory version of the T-maze task by moving the location of the tone cue in order to determine if this affects the strength of the restructuring seen in the firing patterns of the striatum as learning progressed. Lastly, we examined the relationship the training induced pattern had to learning the tone-turn association and to other things that changed systematically throughout learning, such as speed. / / by Terra D. Barnes. / Ph.D.

Thesis (Ph.D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 1998. / Includes bibliographical references (p. 163-168). / The goal of this thesis is to explore the contributions of heredity and environment to language development using the twin method. This method consists of comparing monozygotic (MZ) twins to dizygotic (DZ) twins to see if MZs have a more similar course of development. Two major aspects of language development are examined: vocabulary and syntax. The development of vocabulary in 43 MZ and 33 same-sex DZ English-speaking twins was studied via parent report. The rate and content of the first 100 or so words in children's productive vocabularies were examined. MZ twins had more similar ages at reaching several milestones of vocabulary growth, more similar overall rates of vocabulary growth, more similar spurts in vocabulary growth, and more similarity in actual words and categories of words they produced than DZ twins, suggesting heritability in early word learning. However, the effects of environment were also prominent in vocabulary development, dwarfing effects of heritability. The development of first word combinations and of first over regularizations of the past tense rule "add -ed'' (e.g., goed for went) were also studied by parent report to examine the contributions of heredity and environment to the development of syntax. Both the age of producing first word combinations and the rate of producing combinations were much more similar in MZ than in DZ twins (N = 24 MZ, 23 DZ). This effect showed large heritability and small effects of environment. The rate of producing past tense over regularizations also appeared to be more similar in MZ than in DZ twins, though this result is tentative due to the small number of twins in this study ( 16 MZ, 11 DZ). Finally, the development of verb tense and agreement was studied longitudinally in 4 MZ and 4 DZ twin pairs. The percentage of correct use of tense/agreement morphemes over time was on average more correlated in the MZ than in the DZ twins, suggesting that differences in the development of the Inflectional system also have genetic influence. / by Jennifer B. Ganger. / Ph.D.

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2019 / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 165-186). / Non-convex optimization and learning play an important role in data science and machine learning, yet so far they still elude our understanding in many aspects. In this thesis, I study two important aspects of non-convex optimization and learning: Riemannian optimization and deep neural networks. In the first part, I develop iteration complexity analysis for Riemannian optimization, i.e., optimization problems defined on Riemannian manifolds. Through bounding the distortion introduced by the metric curvature, iteration complexity of Riemannian (stochastic) gradient descent methods is derived. I also show that some fast first-order methods in Euclidean space, such as Nesterov's accelerated gradient descent (AGD) and stochastic variance reduced gradient (SVRG), have Riemannian counterparts that are also fast under certain conditions. In the second part, I challenge two common practices in deep learning, namely empirical risk minimization (ERM) and normalization. Specifically, I show (1) training on convex combinations of samples improves model robustness and generalization, and (2) a good initialization is sufficient for training deep residual networks without normalization. The method in (1), called mixup, is motivated by a data-dependent Lipschitzness regularization of the network. The method in (2), called Zerolnit, makes the network update scale invariant to its depth at initialization. / by Hongyi Zhang. / 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. "February 2019." / Includes bibliographical references (pages 97-103). / Despite rapid recent progress in machine perception and models of biological perception, fundamental questions remain open. In particular, the paradigm underlying these advances, pattern recognition, requires large amounts of training data and struggles to generalize to situations outside the domain of training. In this thesis, I focus on a broad class of perceptual concepts - those that are generated by the composition of multiple causal processes, in this case certain physical interactions - that human use essentially and effortlessly in making sense of the world, but for which any specific instance is extremely rare in our experience. Pattern recognition, or any strongly learning-based approach, might then be an inappropriate way to understand people's perceptual inferences. / I propose an alternative approach, compositional simulation, that can in principle account for these inferences, and I show in practice that it provides both qualitative and quantitative explanatory value for several experimental settings. Consider a box and a number of marbles in the box, and imagine trying to guess how many there are based on the sound produced when the box is shaken. I demonstrate that human observers are quite good at this task, even for subtle numerical differences. Compositional simulation hypothesizes that people succeed by leveraging internal causal models: they simulate the physical collisions that would result from shaking the box (in a particular way), and what those collisions would sound like, for different numbers of marbles. They then compare their simulated sounds with the sound they heard. / Crucially these simulation models can generalize to a wide range of percepts, even those never before experienced, by exploiting the compositional structure of the causal processes being modeled, in terms of objects and their interactions, and physical dynamics and auditory events. Because the motion of the box is a key ingredient in physical simulation, I hypothesize that people can take cues to motion into account in our task; I give evidence that people do. I also consider the domain of unfamiliar objects covered by cloth. a similar mechanism should enable successful recognition even for unfamiliar covered objects (like airplanes). I show that people can succeed in the recognition task, even when the shape of the object is very different when covered. Finally, I show how compositional simulation provides a way to "glue together" the data received by perception (images and sounds) with the contents of cognition (objects). / I apply compositional simulation to two cognitive domains: children's intuitive exploration (obtaining quantitative prediction of exploration time), and causal inference from audiovisual information. / by Max Harmon Siegel. / 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. "February 2019." / Includes bibliographical references (pages 120-153). / Language is the primary means for human interaction, and communicative success requires an ability to reason about a conversation partner's beliefs, desires, and goals (e.g., Grice 1957, 1968, 1975; Sperber & Wilson, 1986). Many questions remain about the precise nature of the relationship between language processing and Theory of Mind (ToM), and their neural substrates. On the one hand, the two domains share an intimate connection given that a) language appears to have been shaped by communicative pressures; b) pragmatic inference is prevalent in language; c) aspects of ToM are important for language acquisition; and d) language is, in turn, critical for acquiring some of the more complex aspects of ToM. But interestingly, prior neuroimaging and patient work have provided some evidence for a dissociation between basic linguistic processing and reasoning about mental states. In this thesis, I use functional MRI to elucidate the relationship between these two core human capacities. / In Chapter 1, I review some of the past research that informs the relationship between language and ToM. In Chapter 2, I examine the topography of language processing and verbal vs. non-verbal ToM in two fMRI datasets (n=90 and n=47). I find that verbal, but not non-verbal, ToM elicits robust responses throughout the language network. Lack of response to non-verbal ToM argues against the core role of the language regions in social reasoning in adulthood. Next, in Chapter 3, I investigate the relationship between the language and ToM networks using a functional correlation approach during several naturalistic conditions (n=55 participants across three studies). I find that although the two networks are dissociable, they also show reliable synchronization both at rest and during story comprehension. This synchronization may provide one mechanism for inter-network information sharing. / And finally, in Chapter 4, I report a case study of an individual (EG) born without a left temporal lobe (likely, due to perinatal stroke damage). Given that in cases of early left hemisphere damage, language processing usually shifts to the right hemisphere, EG presents an interesting case study of how language and ToM - whose most ToM-selective component resides in the right temporoparietal junction - can co-exist within a single temporal lobe. I show that language and ToM overlap to a greater extent in EG compared to a large normative dataset, and some areas that support ToM in neurotypical individuals instead support language processing in EG. However, both their basic language processing, and pragmatic processing appear fully intact. / These results suggest that a single temporal lobe is sufficient to support both language processing and ToM, and that some cortical areas may assume either ToM or language functions, pointing to a deep relationship between the two systems, and the inter-changeable nature of their cortical substrates at least early in development. / by Alexander Paunov. / Ph. D. / Ph.D. Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences