Dissociated neuronal cell cultures grown on substrate integrated microelectrode arrays (MEAs) generate spontaneous activity that can be recorded for up to several weeks. The signature wave shapes from extracellular recording of neuronal activity display a great variety of shapes with triphasic signals predominating. I characterized extracellular recordings from over 600 neuronal signals. I have preformed a categorical study by dividing wave shapes into two major classes: (type 1) signals in which the large positive peak follows the negative spike, and (type 2) signals in which the large positive peak precedes the negative spike. The former are hypothesized to be active signal propagation that can occur in the axon and possibly in soma or dendrites. The latter are hypothesized to be passive which is generally secluded to soma or dendrites. In order to verify these hypotheses, I pharmacologically targeted ion channels with tetrodotoxin (TTX), tetraethylammonium (TEA), 4-aminopyridine (4-AP), and monensin.

Changes in spontaneous spike activities recorded from murine frontal cortex networks grown on substrate-integrated microelectrodes were used to determine the dissociation constant (KB) of three GABAA antagonists. Neuronal networks were treated with fixed concentrations of GABAA antagonists and titrated with muscimol, a GABAA receptor agonist. Muscimol decreased spike activity in a concentration dependent manner with full efficacy (100% spike inhibition) and a 50% inhibitory concentration (IC50) of 0.14 ± 0.05 µM (mean ± SD, n=6). At 10, 20, 40 and 80 µM bicuculline, the muscimol IC50 values were shifted to 4.3 ± 1.8 µM (n=6), 6.8 ± 1.7 µM (n=6), 19.3 ± 3.54 µM (n=10) and 43.5 µM (n=2), respectively (mean ± SD). Muscimol titration in the presence of 10, 20, 40 µM of gabazine resulted in IC50s values of 20.1 (n=2), 37.17 (n=4), and 120.45 (n=2), respectively. In the presence of 20, 80, and 160 µM of TMPP (trimethylolpropane phosphate) the IC50s were 0.86 (n=2), 3.07 (n=3), 6.67 (n=2) µM, respectively. Increasing concentrations of GABAA antagonists shifted agonist log concentration-response curves to the right with identical efficacies, indicating direct competition for the GABAA receptor. A Schild plot analysis with linear regression resulted in slopes of 1.18 ± 0.18, 1.29 ± 0.23 and 1.05 ± 0.03 for bicuculline, gabazine and TMPP, respectively. The potency of antagonists was determined in terms of pA2 values. The pA2 values were 6.63 (gabazine), 6.21 (bicuculline), and 5.4 (TMPP). This suggests that gabazine has a higher binding affinity to the GABAA receptor than bicuculline and TMPP. Hence, using spike rate data obtained from population responses of spontaneously active neuronal networks, it is possible to determine key pharmacological properties of drug-receptor interactions.

Simple circuits perform simple tasks. Complex circuits can perform more complicated tasks. This is true for artificial circuits and for brain circuits. As is known from artificial networks, a complexity that makes circuits substantially more powerful is distributing learning across multiple layers. In fact, most brain circuits in vertebrate systems are multi-layer circuits (but for few that perform simple reflexes) in which learning is distributed across layers. Despite the crucial contribution of learning in middle layer neurons to the output of the circuits they are embedded in, there is little understanding of the principles defining this contribution. A very common feature in brain circuits is that middle layer neurons generate two types of signals, known as spikes. These middle layer neurons commonly have long dendrites where they generate dendritic spikes. As well, like most neurons, they generate axonal spikes near the cell body. Neurons exhibiting these two spike types include pyramidal cells in the neo-cortex and the hippocampus, the Purkinje cells in the cerebellum and many more. In this thesis I study another circuit that contains middle layer neurons, the electrosensory lateral lobe (ELL) of the mormyrid fish. The ELL is a tractable brain circuit in which the middle layer neurons generate dendritic and axonal spikes. In this thesis I show that these spike types are not two different expressions of the same inputs. Rather, they have a symbiotic relationship. Instead of all inputs triggering both spikes, some inputs can selectively drive dendritic spikes. The dendritic spikes in return modify the synaptic strength of another set of inputs. The modified inputs are then transmitted to downstream neurons via the axonal spikes, which contributes a desired signal to the output of the circuits. Effectively there is a separation of learning and signaling in the middle layer neurons through the two spike types. Having two types of spikes in the same neuron doing different computations enormously expands the computational power of the neuron. But, being in the same neuron means the separation of function is constrained and needs to be supported by biophysical principles. I have thus built a biophysical model to understand the biophysical principles underlying the separation of function. I show that in the middle layer neurons of the ELL, the axonal spikes are strongly reduced in amplitude as they backpropagate to the apical dendrites, yet they remain crucial in driving dendritic spikes. Critically, modulation of inhibitory inputs can selectively dial up or down the ability of the backpropagating axonal spikes to drive dendritic spikes. Thus, a set of inhibitory modulating inputs can selectively modulate dendritic spikes. Having learning in different layers contributing to the outcome of the circuit, naturally leads to asking how the work is divided across layers and neuron types within the circuit. In this thesis I answer this question in the context of the outcome of the ELL circuit. Finally, another signature of a complex circuit is the ability to integrate many different inputs, usually in middle layer neurons, to generate sophisticated outputs. A goal for scientists studying systems neuroscience is to understand how this integration works. In this thesis I provide a coherent model of a learning behavior called vestibulo occular reflex (VOR) adaptation, that depends on the integration of separate inputs to yield a learned behavior. The VOR is a simple reflex generated in the brain stem. Inputs from the brain stem are also sent to an area in the cerebellar cortex called the flocculus. The flocculus also receives another set of inputs that generate a different behavior, called smooth-pursuit. The integration of VOR inputs with smooth-pursuit inputs in the flocculus generate VOR adaptation. Understanding complex circuits is one of the greatest challenges for today's neuroscientists. In this thesis I tackle two such circuits and hope that through a better understandings of these circuits we gain principles that apply to other circuits and thereby advance our understanding of the brain.

The purpose of this thesis was to determine the neuronal circuitry mediating U50,488-induced locomotion in preweanling rats. To this end, preweanling rats received bilateral electrolytic lesions of the ventromedial thalamus or superior colliculus and, two days later, the same rats received a challenge injection of U50,488. It was predicted that bilateral lesions of the ventromedial thalamus or superior colliculus would attenuate the U50,488-induced locomotor activity of 18-day-old rats.

The human brain relies on the complex interactions of distinct brain regions to support cognitive processes. The interplay between the hippocampus and neocortical regions plays a key role in the formation, storage, and retrieval of long-term episodic memories. Oscillatory activities during sleep are a fundamental mechanism that binds distributed neuronal networks into functionally coherent ensembles. However, the large-scale hippocampal-neocortical oscillatory mechanisms that support flexible modulation of long-term memory remain poorly understood. Furthermore, alterations to physiologic spatiotemporal patterns that are essential for intact memory function can result in pathophysiology in brain disorders such as focal epilepsy. Investigating how epileptic network activity disrupts connectivity in distributed networks and the organization of oscillatory activity are additional crucial areas that require further research. Our experiments on rodents and human patients with epilepsy have provided valuable insights into these mechanisms. In rodents, we used high-density microelectrode arrays in tandem with hippocampal probes to analyze intracranial electroencephalography (iEEG) from multiple cortical regions and the hippocampus. We identified key hippocampal-cortical oscillatory biomarkers that were differentially coordinated based on the age, strength, and type of memory. We also analyzed iEEG from patients with focal epilepsy and demonstrated how individualized pattern of pathologic-physiologic coupling can disrupt large-scale memory circuits. Our findings may offer new opportunities for therapies aimed at addressing distributed network dysfunction in various neuropsychiatric disorders.

As theoretical neuroscience has grown as a field, machine learning techniques have played an increasingly important role in the development and evaluation of theories of neural computation. Today, machine learning is used in a variety of neuroscientific contexts from statistical inference to neural network training to normative modeling. This dissertation introduces machine learning techniques for use across the various domains of theoretical neuroscience, and the application of these techniques to build theories of neural circuits. First, we introduce a variety of optimization techniques for normative modeling of neural activity, which were used to evaluate theories of primary motor cortex (M1) and supplementary motor area (SMA). Specifically, neural responses during a cycling task performed by monkeys displayed distinctive dynamical geometries, which motivated hypotheses of how these geometries conferred computational properties necessary for the robust production of cyclic movements. By using normative optimization techniques to predict neural responses encoding muscle activity while ascribing to an “untangled” geometry, we found that minimal tangling was an accurate model of M1. Analyses with trajectory constrained RNNs showed that such an organization of M1 neural activity confers noise robustness, and that minimally “divergent” trajectories in SMA enable the tracking of contextual factors. In the remainder of the dissertation, we focus on the introduction and application of deep generative modeling techniques for theoretical neuroscience. Specifically, both techniques employ recent advancements in approaches to deep generative modeling -- normalizing flows -- to capture complex parametric structure in neural models. The first technique, which is designed for statistical generative models, enables look-up inference in intractable exponential family models. The efficiency of this technique is demonstrated by inferring neural firing rates in a log-gaussian poisson model of spiking responses to drift gratings in primary visual cortex. The second technique is designed for statistical inference in mechanistic models, where the inferred parameter distribution is constrained to produce emergent properties of computation. Once fit, the deep generative model confers analytic tools for quantifying the parametric structure giving rise to emergent properties. This technique was used for novel scientific insight into the nature of neuron-type variability in primary visual cortex and of distinct connectivity regimes of rapid task switching in superior colliculus.

As animals explore and interact with their surroundings, information about their environment is constantly processed from sensory stimuli into perception. This information informs their behavior, decision-making, and understanding of their world. Information processing and perception have long been thought to be modulated by the behavior state of the animal (Cano et al. 2006; Niell and Stryker 2010; Polack et al. 2013; Poulet & Petersen, 2008; Briggs 2013, Schölvinck et al. 2015). Previous research has shown that behavior state strongly correlates with perceptual performance in a sensory discrimination task in rodents (McGinley et al. 2015, Schriver et al., 2018). However, the neural correlates behind this modulation of perception, information processing, and behavioral performance are not yet fully understood. The first part of this work investigates the relationship between cell-type specific spontaneous cortical activity and behavior state as defined by pupil-linked arousal. Spontaneous activity is essential in understanding the link between behavior state and information processing as it serves as the baseline state of activity prior to processing any stimuli information. Within the sensory cortices, excitatory and inhibitory neurons work in unison to dictate network activity. Three main classes of cortical inhibitory neurons are somatostatin-expressing neurons (SST), vasointestinal peptide-expressing neurons (VIP), and parvalbumin-expressing neurons (PV). These four cell types comprise the VIP disinhibitory circuit, in which VIP neurons disinhibit excitatory neurons by inhibiting PV and SST neurons. PV and SST neurons directly inhibit excitatory cells, so by suppressing their activity VIP neurons indirectly disinhibit excitatory cells. This circuit is a vitally important system used to modify excitatory activity in all cortical regions. The spontaneous activity of excitatory neurons and three classes of inhibitory neurons (somatostatin-expressing neurons (SST), vasointestinal peptide-expressing neurons (VIP), and parvalbumin-expressing neurons (PV)) was individually examined in this study. To visualize in-vivo spontaneous cortical activity, a genetically encoded calcium indicator (GCaMP) was expressed in the somatosensory cortex, and the population-level neural activity was imaged using fiber photometry. Despite the relationship between these neurons as defined by the VIP disinhibitory circuit, the spontaneous activity of excitatory, VIP, PV, and SST neurons was found to positively correlate with pupil size for all of these neuron types. This supports the theory that VIP and other interneuron types may be active in various functions, not just the disinhibition of excitatory cells. Pupil-evoked activity, or spontaneous activity during highly aroused states, was also found to positively correlate with pupil size for all cell types and had the strongest correlation for all correlation types. Therefore, pupil-linked arousal level relates to the increased activity of both excitatory and inhibitory cortical cells. While the first chapter focuses on spontaneous activity, the second focuses on stimulus-evoked activity. Stimulus-evoked activity in the somatosensory pathway can be caused by both internally generated stimuli and external stimuli. In the first step of sensory processing, the sensory receptors cannot distinguish between these two types of stimuli. However, the differentiation between the two is necessary in order to distinguish self from non-self. The motor-related timing signals that influence sensory processing and enable distinction between internally generated and external stimuli is termed corollary discharge. Where and how the mechanism of corollary discharge occurs in the somatosensory system is not well understood. To investigate corollary discharge in the somatosensory system, the neural activity in the somatosensory cortex was analyzed during internally generated stimuli and during delivery of external stimuli. More specifically, the activity in the vibrissa somatosensory cortex of rodents during self-induced whisking and during delivery of an air puff to the whiskers was examined. In the primary and secondary somatosensory cortex, excitatory activity was inhibited just prior to whisking and suppressed to a lower level during whisking in comparison to the activity level during air puff delivery. The three main classes of inhibitory neurons were studied to explore the possibility of local inhibition causing this suppression of the excitatory signal during whisking. VIP, PV and SST neurons all exhibited a similar pre-whisking inhibition and suppression of activity during whisking, eliminating the possibility of their role in pre-whisking inhibition and whisking activity suppression. Other regions involved in the somatosensory pathway and sensorimotor processing, such as the thalamus and motor cortex, were also found to not contribute to pre-whisking inhibition or whisking activity suppression as they were also found to exhibit the same phenomenon. After ruling out cortical inhibitory neurons and somatosensory regions in the involvement of corollary discharge, external higher-order regions were investigated. Previous studies on the sources of corollary discharge in the cerebellum have shown corollary discharge signals originate from coordination of several different higher-order brain regions (Person A., 2019). To determine these potential regions for somatosensory corollary discharge, viral tracing vectors were used to locate regions with long-range inhibitory projections to the somatosensory cortex. The globus pallidus (GP) was first investigated due to its role in voluntary movement and projections to the frontal cortex (Saunders et al. 2015). However, no inhibitory projections from the GP to the somatosensory cortex were found. The striatum, which is mainly GABAergic (and therefore inhibitory), also seemed to be a likely candidate. Preliminary tracing results suggest the striatum does have inhibitory projections to the somatosensory cortex. Further studies of both retrograde and anterograde tracing must be performed to confirm this finding. Nonetheless, the evidence of corollary discharge as seen through pre-whisking inhibition and the suppression of activity during whisking in S1, S2, thalamus, and motor cortex is a novel finding and opens up many avenues for further research.

Complex genetic mechanisms both endow developing neuronal subtypes with distinct molecular identities and translate those identities into the signatures of cell surface axon guidance molecules that direct neural circuit assembly. The final steps of this process, where axon guidance molecules instruct circuit outcomes, are well-understood. However, the upstream identity molecules that define guidance molecule signatures, and the molecular mechanisms by which cell type identity is transformed into these signatures, remain enigmatic. The murine olfactory system contains nearly 1,5000 olfactory sensory neuron (OSN) subtypes which are intermixed in the olfactory epithelium (OE). Each OSN subtype expresses a unique olfactory receptor (OR) protein which both tunes its response properties to odorants in the environment and acts as an identity molecule that ensures all axons of a given OSN type converge to a single set of target glomeruli in the olfactory bulb (OB). Using a combination of bioinformatic and mouse genetic approaches, we have discovered an unanticipated role for endoplasmic reticulum stress (ER stress) and the unfolded protein response (UPR) in the translation of OR identity to OSN axon guidance molecule expression and glomerular targeting. We find that slight differences in OR amino acid sequences lead to differential activation of the ER stress sensor PERK in different OSN subtypes. Graded patterns of the UPR are then interpreted through a master regulator transcription factor, Ddit3, which controls a set of stress-responsive axon guidance molecules that orchestrate the process of glomerular segregation in the OB. Our results define a novel paradigm for axon guidance in which graded activation of a canonical stress response pathway is leveraged towards the conversion of discrete neuronal identities into discrete circuit formation outcomes. These findings may be widely relevant for the formation of neural circuits across a variety of systems.

Escape behavior is the critical output of rapid sensorimotor processing in the brain that allows animals to sense danger and avoid it. The circuit structures and mechanisms that underlie escape are still under investigation. Drosophila larvae are an advantageous system for studying the neuromuscular circuitry of escape behavior. When threatened with harmful mechanical touch, heat, or light, larvae perform C-shaped bending and lateral rolling, followed by rapid forward crawling. The sensory input and neural circuitry that promotes escape in the larva have been extensively characterized, but we do not understand how bending and rolling motor programs are generated by the larval neuromuscular system. This work identifies the movement patterns, muscle activities, and motor circuit features that drive escape behavior. High-speed imaging approaches reveal that larvae select between four distinct, interchangeable patterns of escape rolling, and that each pattern consists of synchronous rotations of every segment as the larva rotates. Investigating electron microscopic reconstructions of premotor and motor neurons elucidates premotor to motor connectivity patterns that could underlie sequential muscle activity that circumnavigates the larva and propels synchronous rotation along the whole body. Volumetric Swept Confocally-Aligned Planar Excitation (SCAPE) microscopy uncovers that, unlike larval crawling, a well-studied form of larval locomotion that is driven by bilaterally symmetric peristaltic waves of muscle activity, the muscle activity during bending and rolling occurs in a circumferential sequence that is synchronous along the larva’s segments. Muscles neighboring the dorsal and ventral midlines of the larva demonstrate left-right symmetric activity during rolling, and ventral muscles appear to drive the propulsion. Shifts in magnitude of left-right symmetric activity in midline muscles allow the larva to transition from initial escape bending into escape rolling. Preliminary computational predictions of PMN activities confirm the likely necessity of strong ventral muscle coactivity for driving escape. Probing specific PMNs during rolling demonstrates robustness of circuits controlling escape and requires further investigation, alongside the role that sensory feedback could play in this behavior. Altogether, these data reveal a new circuit organization and motor activity pattern that underlie the coordination of muscles during an escape sequence. Future work could reveal circuit components necessary for escape, including the mechanistic basis for action selection, behavioral maintenance, and behavioral flexibility.

In this work, we develop models of neural circuits and plasticity rules that underlie different forms of learning and memory, with a focus on learning processes that involve multiple brain regions. We begin by surveying the literature on synaptic plasticity rules and implementations of learning algorithms in the brain. Each subsequent chapter presents a model of how a specific aspect of learning is implemented biologically, based on experimental evidence and normative considerations. We first focus on the neural basis of reinforcement learning in the basal ganglia. We show that in order to enable effective learning when control of behavior is distributed across multiple regions (``off-policy reinforcement learning''), classic models of dopamine activity must be adapted to include an additional action-sensitive component. We also show that the known plasticity rules of direct and indirect-pathway striatal projection neurons are inconsistent with existing models of striatal codes for action. We propose and find experimental support for a new model of striatal activity driven by efferent input. This model is functionally compatible with striatal plasticity rules and enables simultaneous multiplexing of action-selection and learning signals, a necessary ingredient for off-policy reinforcement learning. We next use an off-policy reinforcement learning model to explain a new experimental finding about the conditions under which learned motor skills are consolidated to be driven by the dorsolateral striatum in rats. We then shift our focus to consider consolidation more broadly, proposing a general model of the advantages of systems in which memories and learned behaviors are consolidated from short-term to long-term learning pathways. In particular, our model proposes that such architectures enable selective filtering of the set of experiences used for learning, which can be essential in noisy environments with many extraneous stimuli. In the appendices, we explore other factors relevant to learning algorithms, including the interaction between multiple sensory modalities, and the problem of credit assignment in multi-layer neural networks. In summary, this work presents a varied set of models of different forms of learning in the brain, emphasizing the cooperative role of plasticity rules and multi-regional circuit architecture in producing functionally useful synaptic weight updates.