arXiv subjects

Shi Gu

Publications and source records attributed to Shi Gu.

At least 37 records · Page 2Linked to original sources

arXiv2021

BRECQ: Pushing the Limit of Post-Training Quantization by Block Reconstruction

We study the challenging task of neural network quantization without end-to-end retraining, called Post-training Quantization (PTQ). PTQ usually requires a small subset of training data but produces less powerful quantized models than Quantization-Aware Training (QAT). In this work, we propose a novel PTQ framework, dubbed BRECQ, which pushes the limits of bitwidth in PTQ down to INT2 for the first time. BRECQ leverages the basic building blocks in neural networks and reconstructs them one-by-one. In a comprehensive theoretical study of the second-order error, we show that BRECQ achieves a good balance between cross-layer dependency and generalization error. To further employ the power of quantization, the mixed precision technique is incorporated in our framework by approximating the inter-layer and intra-layer sensitivity. Extensive experiments on various handcrafted and searched neural architectures are conducted for both image classification and object detection tasks. And for the first time we prove that, without bells and whistles, PTQ can attain 4-bit ResNet and MobileNetV2 comparable with QAT and enjoy 240 times faster production of quantized models. Codes are available at https://github.com/yhhhli/BRECQ.

cs.LG↗

arXiv2021

A Free Lunch From ANN: Towards Efficient, Accurate Spiking Neural Networks Calibration

Spiking Neural Network (SNN) has been recognized as one of the next generation of neural networks. Conventionally, SNN can be converted from a pre-trained ANN by only replacing the ReLU activation to spike activation while keeping the parameters intact. Perhaps surprisingly, in this work we show that a proper way to calibrate the parameters during the conversion of ANN to SNN can bring significant improvements. We introduce SNN Calibration, a cheap but extraordinarily effective method by leveraging the knowledge within a pre-trained Artificial Neural Network (ANN). Starting by analyzing the conversion error and its propagation through layers theoretically, we propose the calibration algorithm that can correct the error layer-by-layer. The calibration only takes a handful number of training data and several minutes to finish. Moreover, our calibration algorithm can produce SNN with state-of-the-art architecture on the large-scale ImageNet dataset, including MobileNet and RegNet. Extensive experiments demonstrate the effectiveness and efficiency of our algorithm. For example, our advanced pipeline can increase up to 69% top-1 accuracy when converting MobileNet on ImageNet compared to baselines. Codes are released at https://github.com/yhhhli/SNN_Calibration.

cs.LG↗

arXiv2021

Optimal Conversion of Conventional Artificial Neural Networks to Spiking Neural Networks

Spiking neural networks (SNNs) are biology-inspired artificial neural networks (ANNs) that comprise of spiking neurons to process asynchronous discrete signals. While more efficient in power consumption and inference speed on the neuromorphic hardware, SNNs are usually difficult to train directly from scratch with spikes due to the discreteness. As an alternative, many efforts have been devoted to converting conventional ANNs into SNNs by copying the weights from ANNs and adjusting the spiking threshold potential of neurons in SNNs. Researchers have designed new SNN architectures and conversion algorithms to diminish the conversion error. However, an effective conversion should address the difference between the SNN and ANN architectures with an efficient approximation \DSK{of} the loss function, which is missing in the field. In this work, we analyze the conversion error by recursive reduction to layer-wise summation and propose a novel strategic pipeline that transfers the weights to the target SNN by combining threshold balance and soft-reset mechanisms. This pipeline enables almost no accuracy loss between the converted SNNs and conventional ANNs with only $\sim1/10$ of the typical SNN simulation time. Our method is promising to get implanted onto embedded platforms with better support of SNNs with limited energy and memory.

cs.NE↗

arXiv2020

Controllability Analysis of Functional Brain Networks

Network control theory has recently emerged as a promising approach for understanding brain function and dynamics. By operationalizing notions of control theory for brain networks, it offers a fundamental explanation for how brain dynamics may be regulated by structural connectivity. While powerful, the approach does not currently consider other non-structural explanations of brain dynamics. Here we extend the analysis of network controllability by formalizing the evolution of neural signals as a function of effective inter-regional coupling and pairwise signal covariance. We find that functional controllability characterizes a region's impact on the capacity for the whole system to shift between states, and significantly predicts individual difference in performance on cognitively demanding tasks including those task working memory, language, and emotional intelligence. When comparing measurements from functional and structural controllability, we observed consistent relations between average and modal controllability, supporting prior work. In the same comparison, we also observed distinct relations between controllability and synchronizability, reflecting the additional information obtained from functional signals. Our work suggests that network control theory can serve as a systematic analysis tool to understand the energetics of brain state transitions, associated cognitive processes, and subsequent behaviors.

q-bio.QM↗

arXiv2019

Unifying Modular and Core-Periphery Structure in Functional Brain Networks over Development

At rest, human brain functional networks display striking modular architecture in which coherent clusters of brain regions are activated. The modular account of brain function is pervasive, reliable, and reproducible. Yet, a complementary perspective posits a core-periphery or rich-club account of brain function, where hubs are densely interconnected with one another, allowing for integrative processing. Unifying these two perspectives has remained difficult due to the fact that the methodological tools to identify modules are entirely distinct from the methodological tools to identify core-periphery structure. Here we leverage a recently-developed model-based approach -- the weighted stochastic block model -- that simultaneously uncovers modular and core-periphery structure, and we apply it to fMRI data acquired at rest in 872 youth of the Philadelphia Neurodevelopmental Cohort. We demonstrate that functional brain networks display rich meso-scale organization beyond that sought by modularity maximization techniques. Moreover, we show that this meso-scale organization changes appreciably over the course of neurodevelopment, and that individual differences in this organization predict individual differences in cognition more accurately than module organization alone. Broadly, our study provides a unified assessment of modular and core-periphery structure in functional brain networks, providing novel insights into their development and implications for behavior.

q-bio.QM↗

arXiv2018

On Structural Controllability of Symmetric (Brain) Networks

The question of controllability of natural and man-made network systems has recently received considerable attention. In the context of the human brain, the study of controllability may not only shed light into the organization and function of different neural circuits, but also inform the design and implementation of minimally invasive yet effective intervention protocols to treat neurological disorders. While the characterization of brain controllability is still in its infancy, some results have recently appeared and given rise to scientific debate. Among these, [1] has numerically shown that a class of brain networks constructed from DSI/DTI imaging data are controllable from one brain region. That is, a single brain region is theoretically capable of moving the whole brain network towards any desired target state. In this note we provide evidence supporting controllability of brain networks from a single region as discussed in [1], thus contradicting the main conclusion and methods developed in [2].

math.OC↗

arXiv2017

Developmental increases in white matter network controllability support a growing diversity of brain dynamics

As the human brain develops, it increasingly supports coordinated control of neural activity. The mechanism by which white matter evolves to support this coordination is not well understood. We use a network representation of diffusion imaging data from 882 youth ages 8 to 22 to show that white matter connectivity becomes increasingly optimized for a diverse range of predicted dynamics in development. Notably, stable controllers in subcortical areas are negatively related to cognitive performance. Investigating structural mechanisms supporting these changes, we simulate network evolution with a set of growth rules. We find that all brain networks are structured in a manner highly optimized for network control, with distinct control mechanisms predicted in child versus older youth. We demonstrate that our results cannot be simply explained by changes in network modularity. This work reveals a possible mechanism of human brain development that preferentially optimizes dynamic network control over static network architecture.

q-bio.NC↗

arXiv2017

Benchmarking measures of network controllability on canonical graph models

Many real-world systems are composed of many individual components that interact with one another in a complex pattern to produce diverse behaviors. Understanding how to intervene in these systems to guide behaviors is critically important to facilitate new discoveries and therapies in systems biology and neuroscience. A promising approach to optimizing interventions in complex systems is network control theory, an emerging conceptual framework and associated mathematics to understand how targeted input to nodes in a network system can predictably alter system dynamics. While network control theory is currently being applied to real-world data, the practical performance of these measures on simple networks with pre-specified structure is not well understood. In this study, we benchmark measures of network controllability on canonical graph models, providing an intuition for how control strategy, graph topology, and edge weight distribution mutually depend on one another. Our numerical studies motivate future analytical efforts to gain a mechanistic understanding of the relationship between graph topology and control, as well as efforts to design networks with specific control profiles.

physics.soc-ph↗

arXiv2017

Optimal Trajectories of Brain State Transitions

The complexity of neural dynamics stems in part from the complexity of the underlying anatomy. Yet how the organization of white matter architecture constrains how the brain transitions from one cognitive state to another remains unknown. Here we address this question from a computational perspective by defining a brain state as a pattern of activity across brain regions. Drawing on recent advances in network control theory, we model the underlying mechanisms of brain state transitions as elicited by the collective control of region sets. Specifically, we examine how the brain moves from a specified initial state (characterized by high activity in the default mode) to a specified target state (characterized by high activity in primary sensorimotor cortex) in finite time. Across all state transitions, we observe that the supramarginal gyrus and the inferior parietal lobule consistently acted as efficient, low energy control hubs, consistent with their strong anatomical connections to key input areas of sensorimotor cortex. Importantly, both these and other regions in the fronto-parietal, cingulo-opercular, and attention systems are poised to affect a broad array of state transitions that cannot easily be classified by traditional notions of control common in the engineering literature. This theoretical versatility comes with a vulnerability to injury. In patients with mild traumatic brain injury, we observe a loss of specificity in putative control processes, suggesting greater susceptibility to damage-induced noise in neurophysiological activity. These results offer fundamentally new insights into the mechanisms driving brain state transitions in healthy cognition and their alteration following injury.

q-bio.NC↗

arXiv2016

Functional Hypergraph Uncovers Novel Covariant Structures over Neurodevelopment

Brain development during adolescence is marked by substantial changes in brain structure and function, leading to a stable network topology in adulthood. However, most prior work has examined the data through the lens of brain areas connected to one another in large-scale functional networks. Here, we apply a recently-developed hypergraph approach that treats network connections (edges) rather than brain regions as the unit of interest, allowing us to describe functional network topology from a fundamentally different perspective. Capitalizing on a sample of 780 youth imaged as part of the Philadelphia Neurodevelopmental Cohort, this hypergraph representation of resting-state functional MRI data reveals three distinct classes of sub-networks (hyperedges): clusters, bridges, and stars, which represent spatially distributed, bipartite, and focal architectures, respectively. Cluster hyperedges show a strong resemblance to the functional modules of the brain including somatomotor, visual, default mode, and salience systems. In contrast, star hyperedges represent highly localized subnetworks centered on a small set of regions, and are distributed across the entire cortex. Finally, bridge hyperedges link clusters and stars in a core-periphery organization. Notably, developmental changes within hyperedges are ordered in a similar core-periphery fashion, with the greatest developmental effects occurring in networked hyperedges within the functional core. Taken together, these results emphasize that the network organization of human brain emerges across multiple scales and evolves substantially through the adolescent period.

q-bio.QM↗

arXiv2016

Explicitly Linking Regional Activation and Function Connectivity: Community Structure of Weighted Networks with Continuous Annotation

A major challenge in neuroimaging is understanding the mapping of neurophysiological dynamics onto cognitive functions. Traditionally, these maps have been constructed by examining changes in the activity magnitude of regions related to task performance. Recently, network neuroscience has produced methods to map connectivity patterns among many regions to certain cognitive functions by drawing on tools from network science and graph theory. However, these two different views are rarely addressed simultaneously, largely because few tools exist that account for patterns between nodes while simultaneously considering activation of nodes. We address this gap by solving the problem of community detection on weighted networks with continuous (non-integer) annotations by deriving a generative probabilistic model. This model generates communities whose members connect densely to nodes within their own community, and whose members share similar annotation values. We demonstrate the utility of the model in the context of neuroimaging data gathered during a motor learning paradigm, where edges are task-based functional connectivity and annotations to each node are beta weights from a general linear model that encoded a linear decrease in blood-oxygen-level-dependent signal with practice. Interestingly, we observe that individuals who learn at a faster rate exhibit the greatest dissimilarity between functional connectivity and activation magnitudes, suggesting that activation and functional connectivity are distinct dimensions of neurophysiology that track behavioral change. More generally, the tool that we develop offers an explicit, mathematically principled link between functional activation and functional connectivity, and can readily be applied to a other similar problems in which one set of imaging data offers network data, and a second offers a regional attribute.

q-bio.NC↗

arXiv2016

The Energy Landscape Underpinning Module Dynamics in the Human Brain Connectome

Human brain dynamics can be profitably viewed through the lens of statistical mechanics, where neurophysiological activity evolves around and between local attractors representing preferred mental states. Many physically-inspired models of these dynamics define the state of the brain based on instantaneous measurements of regional activity. Yet, recent work in network neuroscience has provided initial evidence that the brain might also be well-characterized by time-varying states composed of locally coherent activity or functional modules. Here we study this network-based notion of brain state to understand how functional modules dynamically interact with one another to perform cognitive functions. We estimate the functional relationships between regions of interest (ROIs) by fitting a pair-wise maximum entropy model to each ROI's pattern of allegiance to functional modules. Local minima in this model represent attractor states characterized by specific patterns of modular structure. The clustering of local minima highlights three classes of ROIs with similar patterns of allegiance to community states. Visual, attention, sensorimotor, and subcortical ROIs tend to form a single functional community. The remaining ROIs tend to form a putative executive control community or a putative default mode and salience community. We simulate the brain's dynamic transitions between these community states using a Markov Chain Monte Carlo random walk. We observe that simulated transition probabilities between basins resemble empirically observed transitions between community allegiance states in resting state fMRI data. These results collectively offer a view of the brain as a dynamical system that transitions between basins of attraction characterized by coherent activity in small groups of brain regions, and that the strength of these attractors depends on the cognitive computations being performed.

q-bio.NC↗

arXiv2016

Autaptic Connections Shift Network Excitability and Bursting

Network architecture forms a critical constraint on neuronal function. Here we examine the role of structural autapses, when a neuron synapses onto itself, in driving network-wide bursting behavior. Using a simple spiking model of neuronal activity, we study how autaptic connections affect activity patterns, and evaluate if neuronal degree or controllability are significant factors that affect changes in bursting from these autaptic connections. We observed that adding increasing numbers of autaptic connections to excitatory neurons increased the number of spiking events in the network and the number of network-wide bursts, particularly in the portion of the phase space in which excitatory synapses were stronger contributors to bursting behavior than inhibitory synapses. In comparison, autaptic connections to excitatory neurons with high average controllability led to higher burst frequencies than adding the same number of self-looping connections to neurons with high modal controllability. The number of autaptic connections required to induce bursting behavior could be lowered by selectively adding autapses to high degree excitatory neurons. These results suggest a role of autaptic connections in controlling network-wide bursts in diverse cortical and subcortical regions of mammalian brain. Moreover, they open up new avenues for the study of dynamic neurophysiological correlates of structural controllability.

q-bio.NC↗

arXiv2016

The Energy Landscape of Neurophysiological Activity Implicit in Brain Network Structure

A critical mystery in neuroscience lies in determining how anatomical structure impacts the complex functional dynamics of human thought. How does large-scale brain circuitry constrain states of neuronal activity and transitions between those states? We address these questions using a maximum entropy model of brain dynamics informed by white matter tractography. We demonstrate that the most probable brain states -- characterized by minimal energy -- display common activation profiles across brain areas: local spatially-contiguous sets of brain regions reminiscent of cognitive systems are co-activated frequently. The predicted activation rate of these systems is highly correlated with the observed activation rate measured in a separate resting state fMRI data set, validating the utility of the maximum entropy model in describing neurophysiologial dynamics. This approach also offers a formal notion of the energy of activity within a system, and the energy of activity shared between systems. We observe that within- and between-system energies cleanly separate cognitive systems into distinct categories, optimized for differential contributions to integrated v.s. segregated function. These results support the notion that energetic and structural constraints circumscribe brain dynamics, offering novel insights into the roles that cognitive systems play in driving whole-brain activation patterns.

q-bio.NC↗

arXiv2016

Cognitive Control in the Controllable Connectome

Cognition is supported by neurophysiological processes that occur both in local anatomical neighborhoods and in distributed large-scale circuits. Recent evidence from network control theory suggests that white matter pathways linking large-scale brain regions provide a critical substrate constraining the ability of single areas to affect control on those processes. Yet, no direct evidence exists for a relationship between brain network controllability and cognitive control performance. Here, we address this gap by constructing structural brain networks from diffusion tensor imaging data acquired in 125 healthy adult individuals. We define a simplified model of brain dynamics and simulate network control to quantify modal and boundary controllability, which together describe complementary features of a region's theoretically predicted preference to drive the brain into different cognitive states. We observe that individual differences in these control features derived from structural connectivity are significantly correlated with individual differences in cognitive control performance, as measured by a continuous performance attention test, a color/shape switching task, the Stroop inhibition task, and a spatial n-back working memory task. Indeed, control hubs like anterior cingulate are distinguished from default mode and frontal association areas in terms of the relationship between their control properties and individual differences in cognitive function. These results provide the first empirical evidence that network control forms a fundamental mechanism of cognitive control.

q-bio.NC↗

arXiv2016

Optimally controlling the human connectome: the role of network topology

To meet ongoing cognitive demands, the human brain must seamlessly transition from one brain state to another, in the process drawing on different cognitive systems. How does the brain's network of anatomical connections help facilitate such transitions? Which features of this network contribute to making one transition easy and another transition difficult? Here, we address these questions using network control theory. We calculate the optimal input signals to drive the brain to and from states dominated by different cognitive systems. The input signals allow us to assess the contributions made by different brain regions. We show that such contributions, which we measure as energy, are correlated with regions' weighted degrees. We also show that the network communicability, a measure of direct and indirect connectedness between brain regions, predicts the extent to which brain regions compensate when input to another region is suppressed. Finally, we identify optimal states in which the brain should start (and finish) in order to minimize transition energy. We show that the optimal target states display high activity in hub regions, implicating the brain's rich club. Furthermore, when rich club organization is destroyed, the energy cost associated with state transitions increases significantly, demonstrating that it is the richness of brain regions that makes them ideal targets.

q-bio.NC↗

arXiv2016

Stimulation-based control of dynamic brain networks

The ability to modulate brain states using targeted stimulation is increasingly being employed to treat neurological disorders and to enhance human performance. Despite the growing interest in brain stimulation as a form of neuromodulation, much remains unknown about the network-level impact of these focal perturbations. To study the system wide impact of regional stimulation, we employ a data-driven computational model of nonlinear brain dynamics to systematically explore the effects of targeted stimulation. Validating predictions from network control theory, we uncover the relationship between regional controllability and the focal versus global impact of stimulation, and we relate these findings to differences in the underlying network architecture. Finally, by mapping brain regions to cognitive systems, we observe that the default mode system imparts large global change despite being highly constrained by structural connectivity. This work forms an important step towards the development of personalized stimulation protocols for medical treatment or performance enhancement.

q-bio.NC↗

arXiv2014

Controllability of Brain Networks

Cognitive function is driven by dynamic interactions between large-scale neural circuits or networks, enabling behavior. Fundamental principles constraining these dynamic network processes have remained elusive. Here we use network control theory to offer a mechanistic explanation for how the brain moves between cognitive states drawn from the network organization of white matter microstructure. Our results suggest that densely connected areas, particularly in the default mode system, facilitate the movement of the brain to many easily-reachable states. Weakly connected areas, particularly in cognitive control systems, facilitate the movement of the brain to difficult-to-reach states. Areas located on the boundary between network communities, particularly in attentional control systems, facilitate the integration or segregation of diverse cognitive systems. Our results suggest that structural network differences between the cognitive circuits dictate their distinct roles in controlling dynamic trajectories of brain network function.

q-bio.NC↗