Computer Science Theses and Dissertations
Permanent URI for this collectionhttp://hdl.handle.net/1903/2756
Browse
6 results
Search Results
Item A NEUROCOMPUTATIONAL MODEL OF CAUSAL REASONING AND COMPOSITIONAL WORKING MEMORY FOR IMITATION LEARNING(2022) Davis, Gregory P; Reggia, James A; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Although contemporary neural models excel in a surprisingly diverse range of application domains, they struggle to capture several key qualities of human cognition that are considered crucial for human-level artificial intelligence (AI). Some of these qualities, such as compositionality and interpretability, are readily achieved with traditional symbolic programming, leading some researchers to suggest hybrid neuro-symbolic programming as a viable route to human-level AI. However, the cognitive capabilities of biological nervous systems indicate that it should be possible to achieve human-level reasoning in artificial neural networks without the support of non-neural symbolic algorithms. Furthermore, the computational explanatory gap between cognitive and neural algorithms is a major obstacle to understanding the neural basis of cognition, an endeavor that is mutually beneficial to researchers in AI, neuroscience, and cognitive science. A viable approach to bridging this gap involves "programmable neural networks" that learn to store and evaluate symbolic expressions directly in neural memory, such as the recently developed "Neural Virtual Machine" (NVM). While the NVM achieves Turing-complete universal neural programming, its assembly-like programming language makes it difficult to express the complex algorithms and data structures that are common in symbolic AI, limiting its ability to learn human-level cognitive procedures. I present an approach to high-level neural programming that supports human-like reasoning using only biologically-plausible neural computations. First, I introduce a neural model that represents graph-based data structures as systems of dynamical attractor states called attractor graphs. This model serves as a temporally-local compositional working memory that can be controlled via top-down neural gating. Then, I present a programmable neural network called NeuroLISP that learns an interpreter for a subset of Common LISP. NeuroLISP features native support for compositional data structures, scoped variable binding, and a shared memory space in which programs can be modified as data. Empirical experiments demonstrate that NeuroLISP can learn algorithms for multiway tree processing, compositional sequence manipulation, and symbolic unification in first-order logic. Finally, I present NeuroCERIL, a neural model that performs hierarchical causal reasoning for robotic imitation learning and successfully learns a battery of procedural maintenance tasks from human demonstrations. NeuroCERIL implements a cognitively-plausible and computationally-efficient algorithm for hypothetico-deductive reasoning, which combines bottom-up abductive inference with top-down predictive verification. Because the hypothetico-deductive approach is broadly relevant to a variety of cognitive domains, including problem-solving and diagnostic reasoning, NeuroCERIL is a significant step toward human-level cognition in neural networks.Item EFFECTIVENESS OF PROXIMAL POLICY OPTIMIZATION METHODS FOR NEURAL PROGRAM INDUCTION(2020) Lin, Runxing; Reggia, James Dr.; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)The Neural Virtual Machine (NVM) is a novel neurocomputational architecturedesigned to emulate the functionality of a traditional computer. A version of the NVM called NVM-RL supports reinforcement learning based on standard policy gradient methods as a mechanism for performing neural program induction. In this thesis, I modified NVM-RL using one of the most popular reinforcement learning algorithms, proximal policy optimization (PPO). Surprisingly, using PPO with the existing all-or-nothing reward function did not improve its effectiveness. However, I found that PPO did improve the performance of the existing NVM-RL if one instead used a reward function that grants partial credit for incorrect outputs based on how much those incorrect outputs differ from the correct targets. I conclude that, in some situations, PPO can improve the performance of reinforcement learning during program induction, but that this improvement is dependent on the quality of the reward function that is used.Item A latent variable modeling framework for analyzing neural population activity(2018) Whiteway, Matthew; Butts, Daniel A; Applied Mathematics and Scientific Computation; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Neuroscience is entering the age of big data, due to technological advances in electrical and optical recording techniques. Where historically neuroscientists have only been able to record activity from single neurons at a time, recent advances allow the measurement of activity from multiple neurons simultaneously. In fact, this advancement follows a Moore’s Law-style trend, where the number of simultaneously recorded neurons more than doubles every seven years, and it is now common to see simultaneous recordings from hundreds and even thousands of neurons. The consequences of this data revolution for our understanding of brain struc- ture and function cannot be understated. Not only is there opportunity to address old questions in new ways, but more importantly these experimental techniques will allow neuroscientists to address new questions entirely. However, addressing these questions successfully requires the development of a wide range of new data anal- ysis tools. Many of these tools will draw on recent advances in machine learning and statistics, and in particular there has been a push to develop methods that can accurately model the statistical structure of high-dimensional neural activity. In this dissertation I develop a latent variable modeling framework for analyz- ing such high-dimensional neural data. First, I demonstrate how this framework can be used in an unsupervised fashion as an exploratory tool for large datasets. Next, I extend this framework to incorporate nonlinearities in two distinct ways, and show that the resulting models far outperform standard linear models at capturing the structure of neural activity. Finally, I use this framework to develop a new algorithm for decoding neural activity, and use this as a tool to address questions about how information is represented in populations of neurons.Item Action compositionality with focus on neurodevelopmental disorders(2016) Claudino, Leonardo; Aloimonos, Yiannis; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)A central question in motor neuroscience is how the Central Nervous System (CNS) would handle flexibility at the effector level, that is, how the brain would solve the problem coined by Nikolai Bernstein as the “degrees of freedom problem”, or the task of controlling a much larger number of degrees of freedom (dofs) that is often needed to produce behavior. Flexibility is a bless and a curse: while it enables the same body to engage in a virtually infinite number of behaviors, the CNS is left with the job of figuring out the right subset of dofs to use and how to control and coordinate these degrees. Similarly, at the level of perception, the CNS seeks to obtain information pertaining to the action and actors involved based on perceived motion of other people’s dofs. This problem is believed to be solved with a particular dimensionality reduction strategy, where action production would consist of tuning only a few parameters that control and coordinate a small number of motor primitives, and action perception would take place by applying grouping processes that would solve the inverse problem, that is to identify the motor primitives and the corresponding tuning parameters used by an actor. These parameters can encode not only information on the action per se, but also higher-order cognitive cues like body language or emotion. This compositional view of action representation has an obvious parallel with language: we can think of primitives as words and cognitive systems (motor, perceptual) as different languages. Little is known, however, about how words/primitives would be formed from low-level signals measured at each dof. Here we introduce the SB-ST method, a bottom-up approach to find full-body postural primitives as a set of key postures, that is, vectors corresponding to key relationships among dofs (such as joint rotations) which we call spatial basis (SB) and second, we impose a parametric model to the spatio-temporal (ST) profiles of each SB vector. We showcase the method by applying SB vectors and ST parameters to study vertical jumps of young adults (YAD) typically developing (TD) children and children with Developmental Coordination Disorder (DCD) obtained with motion capture. We also go over a number of other topics related with compositionality: we introduce a top-down system of tool-use primitives based on kinematic events between body parts and objects. The kinematic basis of these events is inspired by the hand-to-object velocity signature reported by movement psychologists in the 1980’s. We discuss the need for custom-made movement measurement strategies to study action primitives on some target populations, for example infants. Having the right tools to record infant movement would be of help, for example, to research in Autism Spectrum Disorder (ASD) where early sensorimotor abnormalities were shown to be linked to future diagnoses of ASD and the development of the typical social traits ASD is mostly known for. We continue the discussion on infant movement measurement where we present an alternative way of processing movement data by using textual descriptions as re- placements to the actual movement signals observed in infant behavioral trials. We explore the fact that these clinical descriptions are freely available as a byproduct of the diagnosis process itself. A typical/atypical classification experiment shows that, at the level of sentences, traditionally used text features in Natural Language Processing such as term frequencies and TF-IDF computed from unigrams and bigrams can be potentially helpful. In the end, we sketch a conceptual, compositional model of action generation based on exploratory results on the jump data, according to which the presence of disorders would be related not to differences in key postures, but in how they are controlled throughout execution. We next discuss the nature of action and actor information representation by analyzing a second dataset with arm-only data (bi-manual coordination and object manipulations) with more target populations than in the jump dataset: TD and DCD children, YAD and seniors with and without Parkinson’s Disease (PD). Multiple group analyses on dofs coupled with explained variances at SB representations suggest that the cost of representing a task as performed by an actor may be equivalent to the cost of representing the task alone. Plus, group discriminating information appears to be more compressed than task-only discriminating information, and because this compression happens at the top spatial bases, we conjecture that groups may be recognized faster than tasks.Item PROCESSING INFORMATION ON INTERMEDIATE TIMESCALES WITHIN RECURRENT NEURAL NETWORKS(2016) Rourke, Oliver; Butts, Dan A; Applied Mathematics and Scientific Computation; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)The cerebral cortex has remarkable computational abilities; it is able to solve prob- lems which remain beyond the most advanced man-made systems. The complexity arises due to the structure of the neural network which controls how the neurons interact. One surprising fact about this network is the dominance of ‘recurrent’ and ‘feedback’ connections. For example, only 5-10% of connections into the earliest stage of visual processing are ‘feedforward’, in that they carry information from the eyes (via the Lateral Geniculate Nucleus). One possible reason for these connec- tions is that they allow for information to be preserved within the network; the underlying ‘causes’ of sensory stimuli usually persist for much longer than the time scales of neural processing, and so understanding them requires continued aggrega- tion of information within the sensory cortices. In this dissertation, I investigate several models of such sensory processing via recurrent connections. I introduce the transient attractor network, which depends on recurrent plastic connectivity, and demonstrate in simulations how it might be involved in the processes of short term memory, signal de-noising, and temporal coherence analysis. I then show how a certain recurrent network structure might allow for transient associative learning to occur on the timescales of seconds using presynaptic facilitation. Finally, I consider how auditory scene analysis might occur through ‘gamma partitioning’. This process uses recurrent excitatory and inhibitory connections to preserve information within the neural network about its recent state, allowing for the separation of auditory sources into different perceptual cycles.Item Sparse Signal Representation in Digital and Biological Systems(2016) Guay, Matthew; Czaja, Wojciech; Applied Mathematics and Scientific Computation; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Theories of sparse signal representation, wherein a signal is decomposed as the sum of a small number of constituent elements, play increasing roles in both mathematical signal processing and neuroscience. This happens despite the differences between signal models in the two domains. After reviewing preliminary material on sparse signal models, I use work on compressed sensing for the electron tomography of biological structures as a target for exploring the efficacy of sparse signal reconstruction in a challenging application domain. My research in this area addresses a topic of keen interest to the biological microscopy community, and has resulted in the development of tomographic reconstruction software which is competitive with the state of the art in its field. Moving from the linear signal domain into the nonlinear dynamics of neural encoding, I explain the sparse coding hypothesis in neuroscience and its relationship with olfaction in locusts. I implement a numerical ODE model of the activity of neural populations responsible for sparse odor coding in locusts as part of a project involving offset spiking in the Kenyon cells. I also explain the validation procedures we have devised to help assess the model's similarity to the biology. The thesis concludes with the development of a new, simplified model of locust olfactory network activity, which seeks with some success to explain statistical properties of the sparse coding processes carried out in the network.