Neuroscience

To survive in natural environments, animals must integrate sensory cues and internal states to generate contextually appropriate behaviors. We leverage the fruit fly Drosophila melanogaster to gain insight into both the strategies and neural codes accompanying this flexibility. In our first study, we examine how male flies choose to either persist with or terminate courtship in response to a threat cue. During this “stay-switch” decision-making, we show that male flies structure their ongoing decisions based on changing social feedback from the female, and that neural dynamics in multiple brain regions predict the animal’s decision outcome. This study reveals that flies are capable of flexible decision-making and reveals that distributed circuits participate in this process. In our second study, we investigate the prevalence of state-dependent neural coding in the fly brain during naturalistic courtship. We show that neural coding of both female-related and self-motion information is widespread in the brain; that coding for female-related cues is enhanced during high-engagement states; and that coding for self-motion information is suppressed during theses states. This study highlights the power of coupling brain-wide imaging with natural behavior to uncover flexible coding properties in the brain. In our third study, we provide a joint-statistical and experimental framework for uncovering the causal interaction map of the fly brain, which we call the “effectome.” We first propose that optogenetic stimulation can be used as an instrumental variable (IV) to distinguish between causal and correlational relationships between neurons. Leveraging the recently completed whole-brain connectome for Drosophila, we show that dynamics in the fly brain consist of a large number of independent activity modes arising from sparse populations of neurons. We show how this organization is amenable to efficiently uncovering the true effectome using our IV approach—a major step beyond correlation-based descriptions of neural activity and small-scale functional mapping. Altogether, our work couples brain-scale analysis of connectivity and dynamics with naturalistic paradigms to uncover fundamental principles of state-dependent coding and behavior.

Training animals to perform cognitive tasks is a critical component of neuroscience research as it enables detailed behavioral and neurophysiological study under controlled conditions. Many tasks require animals to withhold movement during stimulus presentation to clearly signal attention and minimize motion-related artifacts in neural recordings. In rodent studies, this fixation is typically accomplished by maintaining a nose-poke in a central port throughout the stimulus presentation. If an animal prematurely terminates fixation before the response cue, the trial is considered a “violation”. Typically, these events are excluded from all analyses, yet they account for a significant fraction of trials and are correlated with worse performance. This dissertation presents a detailed analysis of these violation events and proposes novel training strategies to mitigate them. In the first chapter, I extend a traditional binary decision-making model to explicitly incorporate violations as a distinct behavioral outcome. I find that past violation history predicts future violations. Moreover, these events negatively interact with stimulus sensitivity, thereby degrading task performance. Building upon these insights, in the second chapter I develop a novel trainingcurriculum, FixGrower, that is designed to efficiently and effectively shape fixation behavior. Compared to legacy training methods, FixGrower significantly reduces training duration while achieving lower violation rates. This curriculum proves robust across various tasks and rodent species, demonstrating its potential as a field-wide standard. Collectively, this work represents a comprehensive characterization of violations and provides optimized methodologies with theoretical support to minimize them. These advancements improve experimental efficiency while enhancing the interpretability and reliability of behavioral and neurophysiological data. Lastly, this work provides a foundation to bridge basic neuroscience with clinical insight to elucidate the mechanisms of impaired motor control that arise in disorders like Parkinson’s and Huntington’s diseases.

Human cognition exhibits a remarkable capacity for learning and generalization from limited experience, extracting abstract principles that enable flexible behavior across diverse contexts. This dissertation investigates the computational and neural mechanisms underlying human-like learning through the lens of meta-learning and inductive biases. The first half of this thesis addresses the challenge of measuring and quantifying abstract knowledge acquisition in both biological and artificial intelligence. I introduce a novel methodology called "task metamers" to behaviorally assess whether learning systems acquire abstract rules. Using this approach, I demonstrate that standard recurrent neural network (RNN) meta-learners fail to capture human abstractions in structured task domains. I then show how augmenting these networks with cognitive representations—natural language and symbolic programs—enables them to better align with human inductive biases. I then further investigate the contributions that both natural language and symbolic representations can make in the development of abstract knowledge that enables generalization. The second half of this thesis investigates the neurocomputational principles that support flexible learning in the brain. Building on meta-reinforcement learning models of the prefrontal cortex-striatum circuit, I explore three key mechanisms: (1) sensory bottlenecks that induce behavioral phenomena like action chunking and neural phenomena like time coding, (2) cortico-striatal loops that enable rapid learning of novel flexible, goal-directed behaviors through recurrent dynamics across multiple brain regions, and (3) functional specialization across cortical areas that emerges from task demands. Together, this work advances our understanding of the powerful, fast, and flexible learning the human brain is capable of, providing both theoretical frameworks and empirical methods for studying human-like learning and generalization.

This thesis traverses a small space in two broad research themes; the first being the striatum’sdopamine responses and how they are involved in initial task learning. The second is joint neural-behavioral statistical models that can and are used to understand both learning and asymptotic performance, and how the brain and behavior influence one another. For the first theme we will study a task learning paradigm where mice engage in a sensory decision making task, and traverse the product space of observations, actions, and outcomes to minimize their thirst (they are water deprived). To this end we provide evidence demonstrating that the initial conditions of the dopamine system in the striatum greatly influence learning outcomes across individual mice. Following this result we will use a reinforcement learning framework to understand how these bonuses lead to differences in learning trajectories between agents. An actor-critic model that makes use of two prediction error signals to reinforce actions neatly explains the main results from the experimental data. For the other theme of this thesis we build and apply a joint neural and behavioral Generalized Linear Model - Hidden Markov Model (GLM-HMM) to asymptotic behavior of the very same data studied in the first theme. There we find that jointly considering both modalities of data changes our statistical descriptions of state-specific behavior and the downstream neural responses occurring in those states.

Understanding how the nervous system flexibly responds to sensory stimuli to generate behav­ior is a major goal in neuroscience. In this thesis, I investigated the mechanosensory circuit of Caenorhabditis elegans to understand how sensory processing is dynamically modulated at the level of behaviors, neuronal circuits, and receptors. I began by investigating the response to a mechanosensory stimulus during turning behavior, as the worm is less likely to reverse in response to a touch stimulus presented during a turn than during a forward behavior. I developed a closed­-loop high­-throughput optogenetic delivery system, capable of delivering optogenetic stimulation to worms triggered on specific behaviors, enabling precise perturbation of neural circuits in behaving animals.I investigated the circuit mechanism underlying turning­ dependent gating of the touch­ evoked reversals. I showed that turning neurons RIV, SMB, and SAA inhibit the reversal neurons upstream of neuron AVA, revealing a mechanism through which turning behavior modulates sensorimotor response. I further identified that heteromeric acetylcholine receptors LGC­-47 and ACC­-1 mediate this context-­dependent mechanosensory response, thus revealing the underlying molecular mecha­nism. I also showed that these heteromeric acetylcholine receptors are involved in the termination of reversal behaviors. I also focused on understanding how opposing sensory stimuli are integrated in the nervous system. To answer this question, I employed the high­throughput optogenetic delivery setup and split­-GAL4 intersectional genetic strategy. When only the anterior touch neurons are activated, a worm responds by slowing down. However, the worm responds by speeding up if only the posterior touch neurons are activated. My investigation showed that the worm’s nervous system linearly integrates the anterior­posterior mechanosensory information to generate motor behaviors. A single­-cell RNA sequencing study identified the tail touch neuron, PLM, as a sexually dimor­phic neuron. My optogenetic interrogation supported this finding by showing that when both the anterior and posterior touch neurons are activated simultaneously, the behavioral response in males and hermaphrodites is significantly different. Together, these studies introduce technological and methodological advances for circuit and molecular level interrogations, and contribute to novel insights into how behavioral and internal states of an animal flexibly gate sensory processing.

Selective attention plays a pivotal role in how we experience and interact with the complex world around us by helping us prioritize the most relevant information at any given moment. Classically, this selective process has been understood as constant and continuous. However, recent evidence suggests that attention alternates rapidly between states of enhanced and diminished perceptual sensitivity. This dissertation explores these temporal dynamics underlying selective attention as well as its ability to guide behavior by being directed externally towards the environment or internally towards working memory. In the first study we develop a task that can detect these fluctuating states of enhanced or diminished perceptual sensitivity at the individual level in a way that is accessible for pediatric and clinical populations. We find that when attending to two features (i.e., colors), attention fluctuates in theta rhythmic cycles creating alternating periods of preference for each feature. In the second study we use this task to investigate these fluctuations in a group of 7 to 10 year olds and find that even the youngest participants display theta rhythmic properties of attention, but that they alternate between states more quickly and experience larger differences between states of enhanced or diminished perceptual sensitivity than adults. The third study extends inquiry into the rhythmic properties of attention to another novel population—adults with schizophrenia. Preliminary results indicate that, similar to the pediatric group, people with schizophrenia differed from adults in the amplitude of their rhythmicity. However, unlike pediatric groups, the adults with schizophrenia showed smaller amplitudes. Furthermore, some data pointed to differences in amplitude being related to symptom severity. In the final study, we investigate how attention can be directed, not only to the external world, but also internally to our memory representations to help guide behavior. We investigate divergent developmental trajectories and explore how the feature complexity of the stimuli being presented impacts the ability to use different spatial attention cues. Collectively, the work presented in this dissertation advances our understanding of the rhythmic temporal dynamics of attention and the interaction between attention and memory.

Chronic stress contributes to detrimental health effects, but our understanding of how stress induces sex differences in brain transcriptomics and behavior is incomplete. In this thesis, I use a combination of behavioral testing, RNA sequencing (RNA-seq), and genetic manipulations in Drosophila melanogaster to explore how social stress reveals latent sex differences in gene expression and stress-relevant behaviors. I begin by introducing mammalian versus insect stress signaling systems and proceed to review methods for stress induction in Drosophila. I also highlight fly behavioral tests for measuring stress-related behaviors, sexual dimorphism in stress responses, and the molecular underpinnings of fly stress behaviors. I then present the effects of social isolation and overcrowding on behavior and survival in Drosophila melanogaster. Finally, I identify sex differences in whole head transcriptomics following social isolation and manipulate the expression of takeout (to) to assess its role in female-specific social isolation-induced increases in sucrose-feeding. Overall, I quantified changes in locomotion, motivated behavior, aggression, and sucrose feeding in flies following social isolation. The most pronounced sex differences were evident in behaviors related to feeding and motivation. RNA profiling revealed isolated female-specific upregulation of over 100 genes, with many of them relating to reproduction and energy metabolism. We manipulated to expression and found that downregulating to in all to-expressing cells decreases sucrose-feeding in isolated females but increases it in isolated males. Upregulating to in all to-expressing cells also decreases sucrose-feeding in isolated females. Altering to expression in neurons did not recapitulate these same effects, suggesting that isolated female-specific to overexpression may occur in non-neuronal cell types in the head. Our results suggest that within a chronic stress context, sex-specific effectors in the head may regulate gene expression related to feeding and macronutrient choice to ensure that females prioritize survival over reproduction. Learning more about this system in flies could provide insight into functionally analogous pathways in humans that may be dysregulated in female-biased stress-related disorders.

In this thesis we consider the use of reinforcement learning (RL) as a formal framework for understanding psychodynamic phenomena. While the field of psychodynamics has influenced a century of theoretical and clinical thinking about psychology, it also faces a number of scientific challenges. We show how RL can provide new explanations for key psychodynamic constructs by grounding them in the principles of reward maximization. The body of the thesis consists of three chapters in which simulations of adaptive agents offer normative accounts of psychodynamic processes. We address the following three phenomena: i) intrapsychic conflict, ii) the dynamics of grief, and iii) the distinct processing of pleasure and pain. We relate these to RL implementations that consider i) modularity, ii) memory replay and iii) reward learning operators, respectively, as central explanatory factors. By providing a computational foundation for psychodynamic concepts, this thesis offers a path toward a more rigorous science of psychological dynamics and their disruption in clinical conditions.

Many fields of science are being revolutionized by new techniques from deep learning to discover dynamics underlying complex high-dimensional data. Adapting these powerful new methods to the challenges of neuroscience, however, is still in its early stages. Here we describe the development of our deep learning method over the years for discovering the nonlinear dynamics of neurons in large populations using only a few latent dimensions. Unlike prominent existing methods, the low-dimensionality of our approach makes the learned dynamical system easier to interpret, even allowing explicit visualization of the vector fields and attractor structures of the system. Our approach leverages neural differential equations (NDEs), a class of deep recurrent neural network models that can implement more complex computations in lower dimensions compared to classical recurrent neural networks. This increased complexity in lower dimensions subsequently helps in extracting interpretable, effective low-dimensional dynamics that may underlie a dataset or task. We address several technical challenges in applying NDEs to neural data, and develop a neural dynamics discovery method called FINDR (Flow-field Inference from Neural Data using deep Recurrent networks) that is built on the backbone of NDEs. To demonstrate its scientific use, we apply FINDR to a variety of neural population datasets, including data from frontal cortex and striatum of rats performing a perceptual decision-making task. We discovered that neural trajectories evolve in two sequential regimes, the first driven by sensory inputs, and the second driven by dynamics internal to the system. The initial regime mediates evidence accumulation, while the subsequent regime subserves decision commitment. This regime transition is coupled to a rapid reorganization in the representation of the decision process in the neural population (a change in the ``neural mode'' along which the process develops). Our results show that the formation of a perceptual choice involves a rapid, coordinated transition in both the dynamical regime and the neural mode of the decision process. Overall, FINDR shows promise as a powerful method for revealing the low-dimensional dynamics of neural populations, and provides a general, interpretable framework for investigating neural computation through the lens of dynamical systems.

The ability to learn from experience is an essential property of cognition and complex behaviour; as such, the question of how learning is implemented in neural circuitry is of paramount importance to the broader field of neuroscience. A major accomplishment in this domain has been the establishment of a link between classic reinforcement learning models and the activity dynamics of various neural populations (chiefly, midbrain dopamine neurons and hippocampal place cells undergoing replay). Recent empirical work employing complex task designs, however, has yielded rich data that challenge the dominant theoretical paradigms in both settings. Within this dissertation, we detail generalisations of the fundamental computational models of dopamine and replay that account for these new data, while retaining explanatory power in the original contexts. Parallel to this, we also present a new estimator for a common statistical model used to analyse learning tasks.

Sensory navigation is crucial for survival and is a common trait across species of all scales. Studying navigation strategies in animals strengthens our understanding of behavioral algorithms and the underlying neural computation. In this thesis, we investigated sensory navigation in the nematode worm C. elegans by combining a novel experimental setup, statistical modeling, neural perturbation and measurements. The integrative approach provides scientific insight into the worm’s navigation strategies, learningdependent behavior, and the underlying neural mechanisms. In the following three chapters, we first built an odor flow chamber and sensory array to control and monitor airborne odor environments, enabling the study of navigation in small animals, such as C. elegans and Drosophila larvae. Secondly, we measured learning-dependent odor-guided navigation in worms and developed a novel statistical model to analyze detailed navigation trajectories. We characterized how learning alters the sensorimotor transformation in worms, demonstrate that the statistical model outperforms classic chemotaxis metric, and revealed distributed neural mechanisms through optogenetic perturbation and genetic ablation. Thirdly, we examined behavioral dynamics along the worm’s chemotaxis trajectory by fitting an extended model with latent states, forming an input-output hidden Markov model for navigation. We found that this state-dependent model better describes the data, reveals complex state-dependent navigation strategies, and pinpoints neural circuits involved in the behavioral dynamics. Additionally, in the last two chapters, we introduced algorithms for stochastic network dynamics that generate state-dependent behavior and analyzed the underlying dynamics. Theoretical projects include a latent variable model for signal progressing in neural imaging, a dynamical model that captures adaptive neural response, and a theoretical framework to analyze nonequilibrium networks. Together, this thesis demonstrates the integration of precise measurements, statistical models, and theory-driven experiments, providing insights into animal navigation algorithms and neural computation.

This thesis presents work investigating how early-life stress (ELS) affects chromatin architecture and development. ELS sensitizes individuals to subsequent stressors, which increases lifetime risk for psychiatric disorders. ELS includes negative experiences such as abuse, neglect, loss of a caregiver, and survival of traumatic events such as exposure to violence or natural disaster. ELS impacts cellular and transcriptomic response to additional stress within the brain’s reward circuitry. How early experiences encode such long-lasting effects is poorly understood but implicates the epigenetic response process, particularly at the level of long-lasting global chromatin architecture remodeling. Here, I describe how epigenetic memory processes could be influenced by ELS, present empirical work showing how chromatin architecture and development change after ELS, and briefly explore potential methods for determining the role individual chromatin regions may play in the ELS susceptibility phenomenon.The empirical work focuses on the effect of ELS on the nucleus accumbens (NAc) — a key limbic brain region, with some discussion of other regions implicated in the reward pathway. Using leading-edge techniques, we find that early-life stress initially opens chromatin in stress-activated cells, suggesting epigenetic priming as a mechanism of stress sensitization. By adulthood, we observe chromatin remodeling throughout the NAc, indicating that the effects of stress are long-lasting and propagate across development into the broader cell population. We also find that deposition of H3K4me1 in NAc during early postnatal development is sufficient to mimic early-life stress and prime hypersensitivity to subsequent stress. We conclude that a molecular memory of early-life stress is encoded at an epigenetic level through changes in chromatin architecture. To complement the initial genome-wide manipulation studies, I have set the groundwork to target specific genomic loci for epigenetic suppression or priming by CRISPR dSaCas9-mediated epigenetic editing, in vitro. This approach was designed to utilize dSaCas9-effector complexes to target enhancers and other regulatory elements to impact chromatin architecture or gene expression, ultimately giving rise to resilience or susceptibility to future stress. While additional work is needed to optimize these systems, this approach has the potential to functionally test whether suppressing or priming specific genomic enhancers during development can promote resilience or sensitivity, respectively, to later stress. Together, this work generates fundamental insights into how stress during key developmental periods increases sensitivity to future stress at an epigenetic level in the brain.

What does it mean to plan? Reinforcement learning theory provides us with a formal definition that identifies the foundational building blocks of planning: the use of an internal action-outcome model to evaluate and guide behavior. The ability to plan comes effortlessly to both animals and humans alike, and it is necessary for successfully navigating the world around us. However, building algorithms that mimic the flexibility of biological planning is exceedingly difficult, and how the brain supports this function is still poorly understood. Identifying the neural mechanisms that underlie efficient planning will provide important insights not only towards computational models of planning, but also for understanding a fundamental function of the brain and why it might fail in the presence of neural disorders. Understanding how the brain plans is the primary goal of my thesis, where I break the problem down into three parts: the “what”, “when”, and “how”. The “what” establishes the problem at hand: what do we know about planning in the brain, and what regions have been implicated? Next, the “when” addresses the way in which we can measure planning and when an animal is most likely to be engaged in planning behavior. Lastly, the “how” begins to answer what neural mechanisms, specifically in dorsal hippocampus, underlie planning behavior. Taken together, the results of this thesis provide new insights into the “when” and “how” the brain plans. We find that we can detect discernible shifts in planning strategy that significantly predict changes in behavioral vigor and neural encoding. We additionally find that the hippocampus represents and utilizes sophisticated state information to bias choice encoding towards upcoming actions dependent on both reward and state transitions.

This thesis investigates the hypothalamic neural circuits controlling social motivation in mice using social operant tasks. Mice, like other animals, show various social behaviors. Controlling motivation for different types of social behaviors are critical for individual survival and reproductive success. However, studying motivations for social behaviors and related neural circuits was often hampered by the difficulty of temporally separating the different phases of a behavior, one with motivation growing and the other with action happening. One solution for this issue is the use of a social operant task, and in the two studies presented here, I fully leverage the social operant task to study the social motivation and its hypothalamic neural basis. In the first study, I investigate the hypothalamic circuits controlling male mice’s aggressive behaviors against subordinate males. In mice, ventromedial hypothalamus ventrolateral part (VMHvl) is known to be a critical hub for aggression, and our team first recorded the neural activity from that region while the recorded male mice engaged in aggressive behaviors, both in a naturalistic setting and in a social operant task, and we confirmed the similarity of neuronal activity pattern during aggression between the two settings. Then we recorded and also optogenetically stimulated the populational activity from its upstream inhibitory hypothalamic regions, and found that the recorded two upstream regions had distinct effects on aggressive motivation and action, as long as the individual neuron’s activity in the downstream VMHvl. Thus, I show how the hypothalamic neural circuits control the aggressive motivation and action in male mice. In the second study, I investigated how the social motivation to different social targets is expressed. Our team built another social operant task in which an animal can choose between male and female animals as the social reward, and manipulated the amount of one of the two social rewards. Our data suggests that male mice have separable, independent social motivations for male and female rewards which female mice do not seem to have. Together these studies increase our understanding of how social motivation is represented andcontrolled in the brain.

A major outstanding challenge in cognitive science is to understand how people perform extrapolations generalizations that are far outside the context of their previous experiences. There has been longstanding controversy regarding the ability of connectionist (i.e., neural- network based) models to attain such behavior in the absence of explicit symbolic primitives. In this thesis, we argue for an alternative perspective that neural networks can learn to behave as if they were implementing a symbolic procedure, provided that they are endowed with certain general inductive biases that shape and constrain the learning process (crucially, these biases do not involve the explicit imposition of any symbolic structure). Specifically, we consider three inductive biases that are deeply motivated by previous work in psychology: The Maximum Entropy Principle, the Relational Bottleneck, and Symmetry. Said in brief, the Maximum Entropy principle prescribes that an agent ought to disperse its internal representations to the fullest possible extent, while Symmetry refers to propensity to attend to repetitive or self-similar structure within the environment, and the Relational Bottleneck refers to a certain architectural constraint which encourages a model to attend exclusively to higher-order relational information at the expense of lower-order perceptual details. In a series of studies, we characterize the effect of multiple combinations of these three learning biases across a wide range of problem domains (analogical reasoning, function learning, and arithmetic). Taken together, our results demonstrate that these biases in combination can effectively enhance the out-of-domain generalization abilities of neural network models, and consequently, due to the psychological plausibility of the inductive biases, point towards a deeper understanding of how such capabilities may be implemented in the human mind.

Optimal sequential decision-making requires considering a vast number of potential action sequences, growing exponentially with every step into the future. The theoretical difficulty of decision-making is compounded by the resource limitations faced by humans and other animals, such as finite time and memory. While these limitations are often cited as the reason for suboptimal behavior, the emerging framework of resource-rationality adopts the perspective that seemingly-suboptimal behavior is actually adaptively tuned to these limitations and makes rational use of limited resources. This dissertation focuses on two broad approaches humans and other animals take to adaptively simplify their decisions. The first is hierarchical representation, intuitively appealing because appropriate choice of hierarchy can decompose complex tasks into simpler subtasks. We develop a resource-rational framework where subgoals are chosen based on how they can simplify the costly process of planning. We identify novel connections between our framework and alternative accounts in simulations, and find that the predictions of our framework are consistent with human behavior in a large-scale behavioral experiment. In a separate study, we run a process-tracing experiment where participants create hierarchically-structured programs and identify a heuristic bias towards reuse that guides hierarchical representations. The second broad approach we consider are heuristic strategies for reinforcement learning tasks. We propose a framework for strategy inference, where strategies are formulated as programs and evaluated based on their task performance in addition to their complexity. Focusing on simulations of bandit problems, we examine how strategies vary based on the weight given to complexity and their relationship to behavioral signatures in previous research.

Early-life stress (ELS) is estimated to affect more than one billion children worldwide every year. ELS includes abuse and neglect, as well as poverty, environmental destruction, and war. ELS poses a major risk for an individual’s future development of depression and other neuropsychiatric disorders, and this early experience may decrease the efficacy of antidepressant drug treatment. Importantly, women are twice as likely to be diagnosed with major depressive disorder compared to men, although less is known about the variation in the efficacy of antidepressant treatment by sex and gender. Understanding the link between ELS and antidepressant outcomes is therefore crucial for improving therapeutic interventions. Here I investigated the predictive role of ELS in altering the efficacy of different classes of antidepressants, focusing on the brain’s nucleus accumbens and sex differences. To emphasize sex differences in ELS, I review in detail the current literature on the role of sex in moderating genome-wide transcriptional patterns in the brain. Further, I conducted translational studies in which I leveraged and performed computational analyses of cross-species genome-wide data from humans and mice treated with different classes of antidepressants. I describe the experiments I conducted of in vivo pharmacological approaches to understand how molecular changes in the brain following ELS may causally alter antidepressant response. Combining computational analyses and experimentation, I found that gene expression patterns in the nucleus accumbens—a brain center of reward and mood—correspond with transcriptional patterns of antidepressant treatment failure. Additionally, transcriptional patterns predicting treatment failure were strongest among female subjects. By profiling global gene expression changes in the brain that contribute to antidepressant treatment outcomes, this research aims to inform translational, clinical decision-making and enhance possible sex-specific efficacy of therapeutic interventions for individuals with depression and a history of ELS.

The hypothesis that midbrain dopamine (DA) neurons broadcast an error for the prediction of reward (reward prediction error, RPE) is among the great successes of computational neuroscience (Houk, Adams, and Barto 1995; Montague, Dayan, and Sejnowski 1996; Schultz, Dayan, and Montague 1997). However, modern empirical results have challenged core aspects of this theory, and my dissertation looks in two such datasets: the first of which delved into how DA responses reflect movement correlates that cannot be explained by a RPE (Parker et al. 2016) and the second of which showed how DA might reflect heterogeneous features rather than convey a scalar, global signal (Engelhard et al. 2019). In my dissertation, I build new models and theories to update the classic RL algorithm TD learning in order to account for these new results. In Chapter 2, I investigate dorsal medial striatum (DMS) projecting DA neurons and how they do not reflect RPE with respect to contralateral movement, but contralateral movement directly and the value of the chosen action. In Chapter 3, I introduce a “feature-specific” PE model in order to explain heterogeneous DA responses in the VTA. I argue that the heterogeneity is a reflection of the input state features upstream of the DA neurons, and show how our model can recapitulate how patterns of heterogeneity might arise for reward prediction errors and movement responses. In Chapter 4, I revisit the dataset from Chapter 2 to determine if the movement correlates could be prediction errors with respect to movements rather than reward. While I could not find an action PE with respect to contralateral movement, I am able to show that DMS-projecting DA population at lever presentation does reflect an identity-free action PE, or an “action-surprise” signal. Overall, my dissertation aims to rethink the PE signal in TD learning, showing that classic RPE accounts of DA need to be updated to account for these new, puzzling data. Together, my models show that DA data may be better understood as a more generalized family of PE models, reflecting a richer signal than just a scalar RPE depending on the inputs they receive.

Some individuals are susceptible to chronic stress, and others are more resilient. While pioneering work has uncovered differences across susceptible and resilient mice after chronic social defeat stress (CSDS), much less is known about the neural signals during stress itself that drive differences in susceptibility. Do the same neural teaching signals implicated in other forms of aversive learning also drive changes during stress, and if so, does their presence before or during the stress predict and produce maladaptive changes? We found that after - but not before - CSDS, the lateral habenula (LHb) is active when susceptible mice are in the proximity of an aggressor strain, but not other strains, consistent with the strain-specific aversion produced by CSDS. During CSDS itself, LHb activity was higher in susceptible mice at the onset of attack by the aggressor, even on the first day of stress. Attack-triggered activation of the LHb during CSDS biased mice towards susceptibility. Brainwide neural activity after stress revealed a striking organization in resilient versus susceptible mice: cortical regions are active in resilient mice, while subcortical regions in susceptible mice. Moreover, LHb activation during CSDS generates sustained and widespread changes in activity that closely resemble the brain-wide correlates of susceptibility. Taken together, our results point to a stronger aversive teaching signal in the LHb from the first day of stress contributing to the formation of the neural and behavioral substrates of susceptibility.

Latent variable models are a powerful tool for analyzing the dynamic structure underlying complex behavior in mammals. In this thesis, I discuss three applications of such mod- els and demonstrate their utility in uncovering novel scientific insights in mice engaging in decision-making and exploratory behaviors. In the first, we use a hidden Markov model with generalized linear model observations (GLM-HMM) to discover that mice occupy multiple internal states, or employ different strategies, when performing a binary decision-making task. This work also reveals that the contribution of the dorsomedial striatum (DMS) to decision-making behavior is state-dependent. In the second application, we extend the use of latent variable models to high-dimensional, continuous-time naturalistic behavior, using a switching Linear Dynamical System (sLDS) to identify discrete states in mice openly ex- ploring their environment. We find that this approach is superior to HMM-based methods in its ability to capture states that correspond to interpretable, consistent behavioral modules. Lastly, we detail novel approaches to improving inference in latent variable models, using ideas from spectral learning theory, with applications to both types of behavior. Overall, this work demonstrates the power of latent variable models to uncover hidden structure and insights into complex behavior in mice, which has broad implications for understanding neural circuits and behavior in both animal models and humans.