Thesis: Ph. D. in Cognitive Science, Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, September, 2020 / Cataloged from student-submitted PDF version of thesis. / Includes bibliographical references (pages 213-224). / The future of machine learning should have a knowledge representation that supports, at a minimum, several features: Expressivity, interpretability, the potential for reuse by both humans and machines, while also enabling sample-efficient generalization. Here we argue that programs-i.e., source code-are a knowledge representation which can contribute to the project of capturing these elements of intelligence. This research direction however requires new program synthesis algorithms which can induce programs solving a range of AI tasks. This program induction challenge confronts two primary obstacles: the space of all programs is infinite, so we need a strong inductive bias or prior to steer us toward the correct programs; and even if we have that prior, effectively searching through the vast combinatorial space of all programs is generally intractable. We introduce algorithms that learn to induce programs, with the goal of addressing these two primary obstacles. Focusing on case studies in vision, computational linguistics, and learning-to-learn, we develop an algorithmic toolkit for learning inductive biases over programs as well as learning to search for programs, drawing on probabilistic, neural, and symbolic methods. Together this toolkit suggests ways in which program induction can contribute to AI, and how we can use learning to improve program synthesis technologies. / by Kevin Ellis. / Ph. D. in Cognitive Science / Ph.D.inCognitiveScience Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences

Thesis: Ph. D. in Neuroscience, Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, September, 2020 / Cataloged from student-submitted PDF version of thesis. / Includes bibliographical references (pages 76-90). / During social interactions, humans and social animals can distinguish not only familiar and novel conspecifics (social recognition) but also between multiple familiar individuals (social specification). Recent studies have implicated hippocampal sub-region dorsal CA2 (dCA2) in social recognition and identified social recognition memory engram in downstream ventral CA1 (vCA1). However, the anatomical site for the storage of social specification memory and its underlying neuroscientific mechanisms are poorly known. Here, we report that social specification memory engrams are stored in vCA1 while social information encoded in dCA2 becomes sharpened as it travels from dCA2 to vCA1 microcircuits within CA2, thereby acquiring a progressive increase in specification through repeating motifs of feed-forward inhibition. Both the inhibition of GABAergic inhibitory neurons in CA2 and reduced activity of excitatory neurons by ablation of oxytocin receptors in the dCA2 to vCA1 microcircuits impair social memory specification. These results suggest that the vCA1 and the multiple feed-forward inhibition motifs in the dCA2 to vCA1 microcircuits are crucial for social memory specification. / by Rosary Yuting Lim. / Ph. D. in Neuroscience / Ph.D.inNeuroscience Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences

Thesis: Ph. D. in Neuroscience, Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, February, 2021 / Cataloged from the official PDF version of thesis. "February 2021." / Includes bibliographical references (pages 190-207). / The biophysical features of neurons are the building blocks of computation in the brain. Dendrites are the physical site of the vast majority of synaptic connections and can expand the information processing capabilities of neurons. Due to their complex morphological attributes and various ion channels, dendrites shape how thousands of inputs are integrated into behaviorally-relevant outputs at the level of individual neurons. However, several long-standing issues limit our understanding of dendritic biophysics. In addition to distorted electrophysiological measurements, prior studies have largely been limited to ex vivo preparations from rodent animal models, providing little insight for computation in the awake human brain. In this thesis, we overcome these limitations to provide new insights on biophysics at the intersection of dendritic morphology and evolution. In chapter 1, we demonstrate that voltage-clamp analysis, which was employed to derive much of our understanding of synaptic transmission, is incompatible with most synapses because they reside on electrically-compartmentalized spines. We also develop new approaches to provide accurate measurements of synaptic strength. Then, in chapter 2, we directly correlate somatic and distal dendritic activity in the awake mouse visual cortex to show an unexpectedly high degree of coupling in vivo. In chapter 3, we perform dendritic recordings in large human neurons to reveal distinct integrative properties from commonly studied rat neurons. Finally, in chapter 4, we characterize neurons in 10 mammalian species to extract evolutionary rules governing neuronal biophysics and uncover human specializations. Together, these four thesis projects expand our understanding of the influence of dendritic geometry and evolution on neuronal biophysics. / by Lou Beaulieu-Laroche. / Ph. D. in Neuroscience / Ph.D.inNeuroscience Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences

Thesis (Ph.D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 1998. / Includes bibliographical references (p. 136-146). / The principles underlying the production of goal directed movements are largely unknown, as evidenced by the lack of artificial systems matching the real-world performance of biological organisms. The ability to transform in real time a vaguely defined goal (skiing downhill) into a detailed motor act (synchronized movements at multiple joints) appropriate under a wide range of previously unencountered circumstances (details of the terrain, snow conditions, etc.) is something we take for granted (after a period of painful learning). This is reminiscent of the subjective ease with which we perceive visually cluttered environments, or follow a conversation in a noisy room. This thesis argues that the classic view of motor control as a unidirectional process, starting with a planning stage ( a learned mapping from goals into actions or a straightforward interpolation connecting a number of intermediate postures into a detailed trajectory) followed by execution of the motor plan, provides an oversimplified and inadequate account of the versatility of everyday human performance. We propose that this standard model of motor control as well as the standard bottom-up models of perception, are inadequate because both perception and motor control are actually "inverse" problems, and are better treated as such. After some theoretical considerations of goal directed multijoint movements, we develop a compu­tational model of sensory-motor processing in intermediate point tasks (including reaching) based on stochastic optimal control theory. In this model planning end execution are not separate stages, but are both integrated into a tight sensory-motor loop which constantly adjusts the ongoing movement to better achieve the specified goal. The model provides a natural account for a number of experimental findings reported here (as well as previously observed phenomena); taken together these experimental results strongly indicate that the motor system updates online its internal estimates of both the envi­ronment and the moving limb, and is always ready to modify its "plan" in favor of a different movement that better achieves the goal under the new circumstances. The observations include eye-hand synchronization patterns, corrections for undetected saccade-triggered visual perturbations and hand inertia anisotropies, effects of desired accuracy on hand kinematics, movement segmentation, speed-curvature relationships. / by Emanuel V. Todorov. / Ph.D.

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, February, 2021 / Cataloged from the official PDF version of thesis. "February 2021." / Includes bibliographical references (pages 119-126). / Recent increases in human longevity have been accompanied by a rise in the incidence of dementia. While a large proportion of aged individuals display pathological hallmarks of neurodegenerative disease, a small number of these individuals are able to maintain healthy cognitive function even in the presence of extensive brain pathology. The molecular mechanisms that govern this neuro-protected state remain unknown, but individuals that exhibit cognitive resilience (CgR) represent a unique source of insight into potential therapies that could preserve brain function in the face of neurodegenerative disease. Here, we employ a two-pronged approach to dissect the mechanism underlying CgR. First, using multiple integrated repositories of clinical and brain transcriptomic data we identified individuals who maintained normal cognition despite harboring a large burden of Alzheimer's disease (AD) pathology. / We observe significant up-regulation of MEF2 family members in these resilient patients when compared to patients whose cognition declined in the presence of pathology. Second, we utilize the only existing animal model of CgR -- environmental enrichment -- / to investigate the molecular mechanisms involved in the induction of resilience. Accessibility of Mef2 binding sites, and expression of Mef2 targets are significantly increased upon enrichment. Additionally, knockdown of Mef2 family members just prior to the initiation of enrichment block its cognitive benefits, demonstrating the necessity of Mef2 activity for achieving the enhanced cognitive potential afforded by enrichment. Neurons lacking Mef2 are hyperexcitable, which is also one of the earliest pathological alterations observed in AD. These results suggest a potential mechanistic link between the Mef2 transcriptional network induced by enrichment and the prevention of disease-associated hyperexcitability. To determine the causal impact of Mef2 on cognition in the context of neurodegeneration, we use a viral approach to manipulate the PS19 mouse model of tauopathy. / Remarkably, in the absence of enrichment, Mef2 overexpression alone is sufficient to improve cognition and reduce hyperexcitability in PS19 mice. Overall, our findings reveal a novel role for MEF2 transcription factors in promoting cognition throughout life, and maintaining cognitive resilience in the context of neurodegenerative disease. / by Scarlett J.V. Barker. / 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, May, 2020 / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 165-183). / The brain codes continuous spatial, temporal, and sensory changes in daily experience. Recent studies suggest the brain also tracks experience as segmented subdivisions (events), but the neural basis for encoding events remains unclear. Here, I present our recent advances to understand the encoding of distinct events at the single cell level. We did preliminary work which revealed distinct neural mechanisms for encoding different spatial contexts. Following this work, we designed a novel maze task for mice which permitted the isolation of neural signals tracking "events" as abstract and discrete entities, separate from sensory changes. This maze task was composed of 4 materially indistinguishable lap events. Using this maze, we reported hippocampal CA1 neurons whose activity was modulated not only by spatial location, but also lap number. These "event-specific rate remapping" (ESR) cells remain lap-specific even when the maze length was unpredictably altered within trials, suggesting ESR cells treated lap events as fundamental units. The activity pattern of ESR cells was reused to represent lap events when the maze geometry was altered from square to circle, suggesting it helped transfer knowledge between experiences. ESR activity was separately manipulable from spatial activity, and may therefore constitute an independent hippocampal code: an "event code" dedicated to organizing experience by events as discrete and transferable units. / by Chen Sun. / Ph. D. / Ph.D. Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences

Thesis: S.M., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, May, 2020 / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 40-47). / Recent advances in the field of optical nanoscopy have equipped molecular neuroscientists with the means to investigate sub-synaptic phenomena with unprecedented spatial resolution. Indeed, early studies employing these methodologies have revealed an extraordinary level of nanoscale spatiotemporal organization of complex and highly dynamic molecular architectures at the pre- and postsynaptic compartments. How exactly these supramolecular complexes are modified through time and experience, though, remains largely unknown. Moreover, the mechanisms by which pre- and postsynaptic protein organization specifies the biochemical and electrophysiological properties of a synapse have yet to be adequately described. The purpose of the current study is thus to develop an inexpensive, accessible platform for super-resolution microscopy (SRM) of synaptic nanostructure. / Though most SRM modalities are too technically and/or financially demanding for widespread use, basic stimulated emission depletion (STED) microscopy systems are now available to researchers at many institutions and require very little training to begin successfully acquiring SR images. One caveat, however, of many such systems is that they may only be capable of SR imaging in two spectral windows due to the inclusion of only one depletion laser in their design. As SR imaging of only two proteins concurrently only promises limited insight into synaptic nanostructure, expansion of these systems' capabilities to localize even one more label would be highly advantageous in these experimental designs and model building. The present study thus develops and optimizes a method to obtain three color SR images on a microscope with a single STED beam. / The unique spectral properties of a long Stokes shift dye, ATTO 490LS, are exploited to add another SR-capable channel without any instrument modifications. This protocol is then applied to image four synaptic proteins simultaneously in primary neuronal culture, with three imaged at STED resolution. More specifically, a tyrosine-phosphorylated subpopulation of the GluA2 glutamate receptor subunit is localized alongside the postsynaptic scaffolding proteins PSD-95 and PSD-93 with the pre-synaptic protein Munc-13 as a confocal landmark. The superior spatial resolutions achieved in this study establish this protocol as an accessible and robust method to introduce an extra SR channel into existing 2-color STED imaging paradigms. / by Mackenzie C. Lee. / S.M. / S.M. Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 1998. / Includes bibliographical references (leaves [135]-[136]). / Two types of processing difficulty are examined by means of electrical recordings taken from the scalp. One type of difficulty seems to be related to syntactic structural anomalies and another is related with memory load due to syntactic complexity. An experiment dealing with structural difficulty reveals the sensitivity of the parser with the argument status of the elements being processed. Memory constraints come into play when processing complex but structurally sound text strings. A number of experiments in this thesis examine a purported metric of complexity, namely, a left anterior negativity. It is argued that the predictive aspects of the parser is responsible for the complexity metric. / by Anthony R. Harris. / Ph.D.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 1998. / Includes bibliographical references (p. 125-146). / This thesis research elucidated principles of somatosensory cortical al organization at the level of the receptive field, cortical map, and cortical area in the rat and human. throughout these studies, a central focus was on the connection between each level of organization and the capacity for cortical reorganization. The rat and human somatoscnsory systems each bring unique opportunities to the study of somatosensory function. Rat SI provides a well-researched system in which detailed, invasive studies can be conducted. Human SI, while less readily available to detailed analysis, is ultimately the relevant preparation for studying the relation between cortical map organization and human perception. In Chapter 1, I examined the convergence of subthreshold (non-action potential evoking) sensory input in individual neuro!\S using whole-cell in vivo recording techniques. Individual neurons in rat SI integrated subthreshold information from an extensive peripheral field, spanning on average 2 vibrissae in each direction from the vibrissa that evoked the largest input (the primary vibrissa). Inputs across the subthreshold receptive field were not homogeneous, as the latency to onset, rise time and prevalence of inhibition varied as a function of the strength of excitatory input and the time poststimulus when they are assessed. These spatial and temporal variables constrain the suprathreshold output of the receptive field, and define the substrate for context-dependent integration of input. In Chapter 2, I described studies of the rat SI vibrissa representation conducted using intrinsic-signal optical imaging. Studies at this level of abstraction PoCG. During tactile stimulation, a precentral gyrus region, corresponding to area 6, and a PoCG region, corresponding to areas 3b,1 and 2, were activated. The borders of these two regions defined an inactive gap in the central sukus region, corresponding to area 3a. Conversely, during proprioceptive/motor stimulation, all three representations were activated. This pattern of activation suggests a strong analogy between the organization of primate and human somatosensory processing streams within the PoCG. Experiment 4 of Chapter 3 presents evidence for the reorganization of human somatosensory cortex following massive deafferentation using perceptual report and fMRI. In agreement with the hypothesis that cortical reorganization is the substrate for phantom perceptions (Ramachandran, 1993; Teuber et al., 1949), two spinal-cord injured (SCI) subjects reported a somatotopic pattern of referred sensations. Further, a subject with phantom perceptions demonstrated greater signal change in the palm representation than an SCI subject without phantom sensations, and normal control subjects (NCS). This finding is in good agreement with increased amplitude in the uncut vibrissa representation following vibrissa trimming/pairing, and suggests that the mechanisms of human cortical reorganization are potentially similar to those observed in rat SI. Taken alone, these studies represent advances in our understanding of the principles of somatosensory cortical organization in rats and humans. In total, these studies provide evidence on three levels of organization, from rat subthreshold receptive fields to human cortical maps, for the importance of dynamic processes in cortical function. / by Christopher I. Moore. / Ph.D.

Thesis: Ph. D. in Neuroscience, Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, September, 2019 / Manuscript. / Includes bibliographical references. / Although much pioneering work has furthered our understanding of the roles the medial geniculate nucleus of the thalamus (MGN) and basolateral amygdala (BLA) play in the formation of associative learning, many open questions remain. MGN is often thought of as only contributing relayed sensory information, and BLA the primary site of importance for the formation of associative learning, aspects of these assumptions have not been directly tested. It is known that both regions are necessary for learning to occur, but the current empirical understanding of the role MGN plays is lacking. Here we present circuit-specific characterization of the information MGN transmits to BLA during discriminative learning, using a combination of optogenetics and in vivo single unit electrophysiology. We demonstrate that while MGN may act as a relay station for information necessary for learning, this is an active process. This input also exhibits dynamic changes between conditions which can also be seen in the BLA population it projects onto. This provides strong evidence suggesting it plays an active role in learning, rather than the assumed role, restricted to sensory relay. Finally, we show that BLA encodes unavoidable punishment associations more strongly than avoidable punishment associations through a shift in the bias of difference encoding of these associations. Taken together, these findings suggest that the assumption of a sensory relay role for MGN in the formation of association central to current dogma is incomplete, and that the avoidability of an associated punishing outcome impacts the magnitude with which BLA encodes the bias between reward and punishment-associated tones. / by Christopher Albert Leppla. / Ph. D. in Neuroscience / Ph. D. in Neuroscience Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences