Center for Complex Systems and Brain Sciences

The midline nuclei of the thalamus, previously characterized as “nonspecific” with undifferentiated connections with the cortex, have been shown to distribute in a specific and highly organized manner to subcortical and cortical structures. The midline thalamus consists of the paraventricular (PV) and paratenial (PT) nuclei, dorsally, and the reuniens (RE) and rhomboid (RH) nuclei, ventrally. The PV and RE nuclei have been examined to a far greater extent than either the PT or RH and have been shown to be involved in various affective and cognitive functions. Generally, PV is more associated with emotional and motivated behaviors such as arousal, feeding, drug addiction, fear, and anxiety, whereas RE is more involved in cognitive and mnemonic functions -- as RE represents a critical bridge between the medial prefrontal cortex (mPFC) and the hippocampal formation.
As afferent projections to PT have not been systemically described, we examined the input to PT comparing it with that to PV, using retrograde anatomical tracer, fluorogold (FG). We found PT and PV are almost exclusively targeted by ‘limbic’ structures of the forebrain. Whereas afferents to PT and PV originate from very similar sources, PT receives stronger input from the cortex and PV from subcortical structures. Notably, PT receives prominent input from the mPFC and orbital (ORB) cortices, two regions associated with cognitive flexibility and decision making.

Working memory (WM) enables the flexible representation of information over short intervals. It is established that WM performance can be enhanced by a retrospective cue presented during storage, yet the neural mechanisms responsible for this benefit are unclear. Here, we tested several explanations for retrospective cue benefits by quantifying changes in spatial WM representations reconstructed from alpha-band (8 - 12 Hz) EEG activity recorded from human participants before and after the presentation of a retrospective cue. This allowed us to track cue-related changes in WM representations with high temporal resolution. Our findings suggest that retrospective cues engage several different mechanisms such as recovery of information previously decreased to baseline after being cued as relevant and protecting the cued item from temporal decay to mitigate information loss during WM storage. Our EEG findings suggest that participants can supplement active memory traces with information from other memory stores. We next sought to better understand these additional store(s) by asking whether they are subject to the same temporal degradation seen in active memory representations during storage. We observed a significant increase in the quality of location representations following a retrocue, but the magnitude of this benefit was linearly and inversely related to the timing of the retrocue such that later cues yielded smaller increases.

Neural network models with many tunable parameters can be trained to approximate functions that transform a source distribution, or dataset, into a target distribution of interest. In contrast to low-parameter models with simple governing equations, the dynamics of transformations learned in deep neural network models are abstract and the correspondence of dynamical structure to predictive function is opaque. Despite their “black box” nature, neural networks converge to functions that implement complex tasks in computer vision, Natural Language Processing (NLP), and the sciences when trained on large quantities of data. Where traditional machine learning approaches rely on clean datasets with appropriate features, sample densities, and label distributions to mitigate unwanted bias, modern Transformer neural networks with self-attention mechanisms use Self-Supervised Learning (SSL) to pretrain on large, unlabeled datasets scraped from the internet without concern for data quality. SSL tasks have been shown to learn functions that match or outperform their supervised learning counterparts in many fields, even without task-specific finetuning. The recent paradigm shift to pretraining large models with massive amounts of unlabeled data has given credibility to the hypothesis that SSL pretraining can produce functions that implement generally intelligent computations.

Although state-of-the-art Convolutional Neural Networks (CNNs) are often viewed as a model of biological object recognition, they lack many computational and architectural motifs that are postulated to contribute to robust perception in biological neural systems. For example, modern CNNs lack lateral connections, which greatly outnumber feed-forward excitatory connections in primary sensory cortical areas and mediate feature-specific competition between neighboring neurons to form robust, sparse representations of sensory stimuli for downstream tasks. In this thesis, I hypothesize that CNN layers equipped with lateral competition better approximate the response characteristics and dynamics of neurons in the mammalian primary visual cortex, leading to increased robustness under noise and/or adversarial attacks relative to current robust CNN layers. To test this hypothesis, I develop a new class of CNNs called LCANets, which simulate recurrent, feature-specific lateral competition between neighboring neurons via a sparse coding model termed the Locally Competitive Algorithm (LCA). I first perform an analysis of the response properties of LCA and show that sparse representations formed by lateral competition more accurately mirror response characteristics of primary visual cortical populations and are more useful for downstream tasks like object recognition than previous sparse CNNs, which approximate competition with winner-take-all mechanisms implemented via thresholding.

Within Deep CNNs there is great excitement over breakthroughs in network performance on benchmark datasets such as ImageNet. Around the world competitive teams work on new ways to innovate and modify existing networks, or create new ones that can reach higher and higher accuracy levels. We believe that this important research must be supplemented with research into the computational dynamics of the networks themselves. We present research into network behavior as it is affected by: variations in the number of filters per layer, pruning filters during and after training, collapsing the weight space of the trained network using a basic quantization, and the effect of Image Size and Input Layer Stride on training time and test accuracy. We provide insights into how the total number of updatable parameters can affect training time and accuracy, and how “time per epoch” and “number of epochs” affect network training time. We conclude with statistically significant models that allow us to predict training time as a function of total number of updatable parameters in the network.

Visual working memory (VWM) is a core cognitive system that supports our ability to maintain and manipulate visual information temporarily when sensory information is no longer present in the environment. VWM and mental rotation, a form of mental imagery, require the ability to generate internal images in the absence of stimuli. Both cognitive processes share visual buffer and are associated with representing and manipulating visual information, however, little is known about the intersection between VWM and mental rotation. In the current work, mental rotation was adopted to study updated mnemonic contents in VWM. In this dissertation, I asked whether the brain mechanisms that support VWM and mental rotation overlap. Participants were asked to remember the orientation of grating or to remember and manipulate, that is mentally rotate, the orientation of grating. Behavioral results showed that mental rotation induced lower fidelity representations of orientation. This confirmed that additional usage in visual buffer to manipulate the visual representation provoked by mental rotation involved negative influence in memory fidelity. In the second study, EEG recording was conducted while participants performed the same task. Visual representations were reconstructed from brain oscillations using the inverted encoding model (IEM). It was found that orientation information from the reconstruction was represented in the amplitude of alpha oscillations (8 – 12 Hz) for both maintained and updated mnemonic contents. Together, this work provides evidence that memory manipulation driven by mental rotation has a decisive effect on the fidelity of visual representations in VWM. Additionally this dissertation demonstrates that the updated memory representations as well as the maintained memory representations are carried in EEG oscillations.

One basic goal of artificial learning systems is the ability to continually learn throughout that system’s lifetime. Transitioning between tasks and re-deploying prior knowledge is thus a desired feature of artificial learning. However, in the deep-learning approaches, the problem of catastrophic forgetting of prior knowledge persists. As a field, we want to solve the catastrophic forgetting problem without requiring exponential computations or time, while demonstrating real-world relevance. This work proposes a novel model which uses an evolutionary algorithm similar to a meta-learning objective, that is fitted with a resource constraint metrics. Four reinforcement learning environments are considered with the shared concept of depth although the collection of environments is multi-modal. This system shows preservation of some knowledge in sequential task learning and protection of catastrophic forgetting in deep neural networks.

The high prevalence of autism spectrum disorder (ASD) results in large costs to individuals, families, and society. Among diagnosed individuals, restrictive and repetitive behaviors (RRBs) correlate with functional impairments substantially impacting wellbeing but remain less studied than social and communication deficits. Brain resting-state functional connectivity (fc) measures intrinsic, potentially RRB-associated neural dynamics. Here, whole-brain (WB), and iterated seed-based (SB)fc guided by the preceding WBfc and a priori hypotheses was performed. Combined results were used to model a brain network beginning with qualitative assessment of its potential functional association with RRBs. Once rigorously defined, the network was used to inform construction of a dynamical systems model of brain activity hypothesized to correlate with RRB severity. Qualitative model behavior tracked expectations of real cortical activity in RRB presentation. Model numerical output was found to correlate with behavioral measures of RRBs to a significantly greater degree than the underlying brain connectivity values themselves did. Some summary measures of model output were also found to correlate significantly, though near threshold, with severity measures in the other two ASD core deficit domains, and particularly, far more extensively than should be expected given the underlying brain connectivity values themselves were apparently effectively wholly uncorrelated with the measures. Significant findings are: (1) dynamical modeling of brain activity can identify significant correlations with symptom manifestation that fc alone cannot; (2) dynamical modeling of brain activity could potentially increase understanding of ASD’s extensive heterogeneity across symptom domains; (3) extensive overlap between the constructed network and known RRB-implicated brain divisions was identified, with cerebellum, increasingly implicated in distributed neocortical functional differences in RRBs in humans and animal models, centrally connected to multiple such divisions; (4) further overlap is found via striatal circuitry, implicated in multiple RRB-like behaviors previously, and forming at least 1/3 of the functional basis for the network’s hypothetical relationship with RRBs; (5) ASD-associated angular gyrus, PFC, ACC overlap was found. This successful tandem application of fc, dynamical modeling, and neurocognitive network theory illustrates the need for broad theoretical approaches in illuminating ASD heterogeneity and the neurocognitive underpinnings of specific ASD presentations.

Even during fixation, the eye is rarely still, as miniature eye movements continue to occur within fixational periods of the eye. These miniature movements are referred to as fixational eye movements. Microsaccades are one of the three types of fixational eye movements that have been identified. Microsaccades have been attributed to different visual processes/phenomena such as fixation stability, perceptual fading, and multistable perception. Still, debates surrounding the functional role of microsaccades in vision ensued, as many of the findings from earlier microsaccade reports contradict one another and the polarity in the field caused by these debates led many to believe that microsaccades do not hold a necessary/specialized role in vision. To gain a deeper understanding of microsaccades and its relevance in vision, we sought out to assess the role of microsaccades in bistable motion perception in our behavioral/eye-tracking study. Observers participated in an eye-tracking experiment where they were asked to complete a motion discrimination task while viewing a bistable apparent motion stimuli. The collected eye-tracking data was then used to train a classification model to predict directions of illusory motion perceived by observers. We found that small changes in gaze position during fixation, occurring within or outside microsaccadic events, predicted the direction of motion pattern imposed by the motion stimuli. Our findings suggest that microsaccades and fixational eye movements are correlated with motion perception and that miniature eye movements occurring during fixation may have relevance in vision.

Both proliferative diabetic retinopathy and exudative age-related macular degeneration are major causes of blindness which are caused by growth of defective, leaky and tortuous blood vessels in the retina. Hypoxia is implicated in triggering both of these diseases and results in induction of HIF-1alpha transcription factor in addition to the angiogenic factor VEGF. Müller cells are the major glial cell in the retina and they contribute to neovascularization in hypoxic regions of the retina through eliciting secretion of growth factors, cytokines and angiogenic factors. As Müller cells span the breadth of the retina they can secrete angiostatic factors as well as neuroprotective trophic factors, the Müller cell is a valuable cell type for targeting by potential new gene therapies. The current investigation tests the hypoxia responsiveness of an AAV vector containing a hybrid hypoxia response element together with a GFAP promoter, and this vector encodes the angiostatic protein decorin, a well characterized multi-receptor tyrosine kinase inhibitor. Decorin may have advantages over other key angiostatic factors such as endostatin or angiostatin by virtue of its multiple anti-angiogenic signaling modalities. We employed Q-RT-PCR to evaluate the cell specificity and hypoxia responsiveness of an AAV-Vector termed AAV-REG-Decorin containing a hybrid HRE and GFAP promoter driving expression of the decorin transgene. The vector also contains a silencer element between the HRE and the GFAP domains to enable low basal expression in normoxia as well as high level inducibility in hypoxia. AAV-REGDecorin was found to elicit high level expression of decorin mRNA in hypoxia with greater than 9 – fold induction of the transgene in hypoxic conditions in astrocytes by comparison to normoxic astrocytes. AAV-REG-Decorin showed low levels of transgene expression by comparison to the positive control vector AAV-CMV -decorin containing the ubiquitously active CMV-promoter. The expression levels of decorin mRNA from AAV-REG-Decorin and from AAV-GFAP-Decorin were low in the PC12 neuronal cell model and in the ARPE19 line of retinal pigment epithelial cells with respect to those of AAV-CMV-decorin and no induction of Decorin mRNA was found with AAV-REGDecorin in these two control cell lines. Our novel gene therapy vector will serve as a platform for testing efficacy in rodent disease models (OIR and laser induced choroidal neovascularization) for assessment of the benefits of tightly regulated antiangiogenic gene therapy eliciting decorin transgene expression, both in terms of timing and the cellular source of production, during the progression of the retinal pathophysiology.