Dense Sample Deep Learning.

Deep learning (DL), a variant of the neural network algorithms originally proposed in the 1980s (Rumelhart et al., 1986), has made surprising progress in artificial intelligence (AI), ranging from language translation, protein folding (Jumper et al., 2021), autonomous cars, and, more recently, human-like language models (chatbots). All that seemed intractable until very recently. Despite the growing use of DL networks, little is understood about the learning mechanisms and representations that make these networks effective across such a diverse range of applications. Part of the answer must be the huge scale of the architecture and, of course, the large scale of the data, since not much has changed since 1986. But the nature of deep learned representations remains largely unknown. Unfortunately, training sets with millions or billions of tokens have unknown combinatorics, and networks with millions or billions of hidden units can't easily be visualized and their mechanisms can't be easily revealed. In this letter, we explore these challenges with a large (1.24 million weights VGG) DL in a novel high-density sample task (five unique tokens with more than 500 exemplars per token), which allows us to more carefully follow the emergence of category structure and feature construction. We use various visualization methods for following the emergence of the classification and the development of the coupling of feature detectors and structures that provide a type of graphical bootstrapping. From these results, we harvest some basic observations of the learning dynamics of DL and propose a new theory of complex feature construction based on our results.

Causally informed activity flow models provide mechanistic insight into network-generated cognitive activations.

Brain activity flow models estimate the movement of task-evoked activity over brain connections to help explain network-generated task functionality. Activity flow models have been shown to accurately generate task-evoked brain activations across a wide variety of brain regions and task conditions. However, these models have had limited explanatory power, given known issues with causal interpretations of the standard functional connectivity measures used to parameterize activity flow models. We show here that functional/effective connectivity (FC) measures grounded in causal principles facilitate mechanistic interpretation of activity flow models. We progress from simple to complex FC measures, with each adding algorithmic details reflecting causal principles. This reflects many neuroscientists' preference for reduced FC measure complexity (to minimize assumptions, minimize compute time, and fully comprehend and easily communicate methodological details), which potentially trades off with causal validity. We start with Pearson correlation (the current field standard) to remain maximally relevant to the field, estimating causal validity across a range of FC measures using simulations and empirical fMRI data. Finally, we apply causal-FC-based activity flow modeling to a dorsolateral prefrontal cortex region (DLPFC), demonstrating distributed causal network mechanisms contributing to its strong activation during a working memory task. Notably, this fully distributed model is able to account for DLPFC working memory effects traditionally thought to rely primarily on within-region (i.e., not distributed) recurrent processes. Together, these results reveal the promise of parameterizing activity flow models using causal FC methods to identify network mechanisms underlying cognitive computations in the human brain.

Latent functional connectivity underlying multiple brain states.

Functional connectivity (FC) studies have predominantly focused on resting state, where ongoing dynamics are thought to reflect the brain's intrinsic network architecture, which is thought to be broadly relevant because it persists across brain states (i.e., is state-general). However, it is unknown whether resting state is the optimal state for measuring intrinsic FC. We propose that latent FC, reflecting shared connectivity patterns across many brain states, better captures state-general intrinsic FC relative to measures derived from resting state alone. We estimated latent FC independently for each connection using leave-one-task-out factor analysis in seven highly distinct task states (24 conditions) and resting state using fMRI data from the Human Connectome Project. Compared with resting-state connectivity, latent FC improves generalization to held-out brain states, better explaining patterns of connectivity and task-evoked activation. We also found that latent connectivity improved prediction of behavior outside the scanner, indexed by the general intelligence factor (g). Our results suggest that FC patterns shared across many brain states, rather than just resting state, better reflect state-general connectivity. This affirms the notion of "intrinsic" brain network architecture as a set of connectivity properties persistent across brain states, providing an updated conceptual and mathematical framework of intrinsic connectivity as a latent factor.

Experimental evidence for a child-to-adolescent switch in human amygdala-prefrontal cortex communication: A cross-sectional pilot study.

Interactions between the amygdala and prefrontal cortex are fundamental to human emotion. Despite the central role of frontoamygdala communication in adult emotional learning and regulation, little is known about how top-down control emerges during human development. In the present cross-sectional pilot study, we experimentally manipulated prefrontal engagement to test its effects on the amygdala during development. Inducing dorsal anterior cingulate cortex (dACC) activation resulted in developmentally-opposite effects on amygdala reactivity during childhood versus adolescence, such that dACC activation was followed by increased amygdala reactivity in childhood but reduced amygdala reactivity in adolescence. Bayesian network analyses revealed an age-related switch between childhood and adolescence in the nature of amygdala connectivity with the dACC and ventromedial PFC (vmPFC). Whereas adolescence was marked by information flow from dACC and vmPFC to amygdala (consistent with that observed in adults), the reverse information flow, from the amygdala to dACC and vmPFC, was dominant in childhood. The age-related switch in information flow suggests a potential shift from bottom-up co-excitatory to top-down regulatory frontoamygdala connectivity and may indicate a profound change in the circuitry supporting maturation of emotional behavior. These findings provide novel insight into the developmental construction of amygdala-cortical connections and implications for the ways in which childhood experiences may influence subsequent prefrontal function.

The Stochastic Delta Rule: Faster and More Accurate Deep Learning Through Adaptive Weight Noise.

Multilayer neural networks have led to remarkable performance on many kinds of benchmark tasks in text, speech, and image processing. Nonlinear parameter estimation in hierarchical models is known to be subject to overfitting and misspecification. One approach to these estimation and related problems (e.g., saddle points, colinearity, feature discovery) is called Dropout. The Dropout algorithm removes hidden units according to a binomial random variable with probability p prior to each update, creating random "shocks" to the network that are averaged over updates (thus creating weight sharing). In this letter, we reestablish an older parameter search method and show that Dropout is a special case of this more general model, stochastic delta rule (SDR), published originally in 1990. Unlike Dropout, SDR redefines each weight in the network as a random variable with mean µwij and standard deviation σwij. Each weight random variable is sampled on each forward activation, consequently creating an exponential number of potential networks with shared weights (accumulated in the mean values). Both parameters are updated according to prediction error, thus resulting in weight noise injections that reflect a local history of prediction error and local model averaging. SDR therefore implements a more sensitive local gradient-dependent simulated annealing per weight converging in the limit to a Bayes optimal network. We run tests on standard benchmarks (CIFAR and ImageNet) using a modified version of DenseNet and show that SDR outperforms standard Dropout in top-5 validation error by approximately 13% with DenseNet-BC 121 on ImageNet and find various validation error improvements in smaller networks. We also show that SDR reaches the same accuracy that Dropout attains in 100 epochs in as few as 40 epochs, as well as improvements in training error by as much as 80%.

Advancing functional connectivity research from association to causation.

Cognition and behavior emerge from brain network interactions, such that investigating causal interactions should be central to the study of brain function. Approaches that characterize statistical associations among neural time series-functional connectivity (FC) methods-are likely a good starting point for estimating brain network interactions. Yet only a subset of FC methods ('effective connectivity') is explicitly designed to infer causal interactions from statistical associations. Here we incorporate best practices from diverse areas of FC research to illustrate how FC methods can be refined to improve inferences about neural mechanisms, with properties of causal neural interactions as a common ontology to facilitate cumulative progress across FC approaches. We further demonstrate how the most common FC measures (correlation and coherence) reduce the set of likely causal models, facilitating causal inferences despite major limitations. Alternative FC measures are suggested to immediately start improving causal inferences beyond these common FC measures.

Attentional Bias in Human Category Learning: The Case of Deep Learning.

Category learning performance is influenced by both the nature of the category's structure and the way category features are processed during learning. Shepard (1964, 1987) showed that stimuli can have structures with features that are statistically uncorrelated (separable) or statistically correlated (integral) within categories. Humans find it much easier to learn categories having separable features, especially when attention to only a subset of relevant features is required, and harder to learn categories having integral features, which require consideration of all of the available features and integration of all the relevant category features satisfying the category rule (Garner, 1974). In contrast to humans, a single hidden layer backpropagation (BP) neural network has been shown to learn both separable and integral categories equally easily, independent of the category rule (Kruschke, 1993). This "failure" to replicate human category performance appeared to be strong evidence that connectionist networks were incapable of modeling human attentional bias. We tested the presumed limitations of attentional bias in networks in two ways: (1) by having networks learn categories with exemplars that have high feature complexity in contrast to the low dimensional stimuli previously used, and (2) by investigating whether a Deep Learning (DL) network, which has demonstrated humanlike performance in many different kinds of tasks (language translation, autonomous driving, etc.), would display human-like attentional bias during category learning. We were able to show a number of interesting results. First, we replicated the failure of BP to differentially process integral and separable category structures when low dimensional stimuli are used (Garner, 1974; Kruschke, 1993). Second, we show that using the same low dimensional stimuli, Deep Learning (DL), unlike BP but similar to humans, learns separable category structures more quickly than integral category structures. Third, we show that even BP can exhibit human like learning differences between integral and separable category structures when high dimensional stimuli (face exemplars) are used. We conclude, after visualizing the hidden unit representations, that DL appears to extend initial learning due to feature development thereby reducing destructive feature competition by incrementally refining feature detectors throughout later layers until a tipping point (in terms of error) is reached resulting in rapid asymptotic learning.

Differences in atypical resting-state effective connectivity distinguish autism from schizophrenia.

Autism and schizophrenia share overlapping genetic etiology, common changes in brain structure and common cognitive deficits. A number of studies using resting state fMRI have shown that machine learning algorithms can distinguish between healthy controls and individuals diagnosed with either autism spectrum disorder or schizophrenia. However, it has not yet been determined whether machine learning algorithms can be used to distinguish between the two disorders. Using a linear support vector machine, we identify features that are most diagnostic for each disorder and successfully use them to classify an independent cohort of subjects. We find both common and divergent connectivity differences largely in the default mode network as well as in salience, and motor networks. Using divergent connectivity differences, we are able to distinguish autistic subjects from those with schizophrenia. Understanding the common and divergent connectivity changes associated with these disorders may provide a framework for understanding their shared cognitive deficits.

Back to the future: The return of cognitive functionalism.

The claims that learning systems must build causal models and provide explanations of their inferences are not new, and advocate a cognitive functionalism for artificial intelligence. This view conflates the relationships between implicit and explicit knowledge representation. We present recent evidence that neural networks do engage in model building, which is implicit, and cannot be dissociated from the learning process.

Modeling Causal Relationship Between Brain Regions Within the Drug-Cue Processing Network in Chronic Cocaine Smokers.

The cues associated with drugs of abuse have an essential role in perpetuating problematic use, yet effective connectivity or the causal interaction between brain regions mediating the processing of drug cues has not been defined. The aim of this fMRI study was to model the causal interaction between brain regions within the drug-cue processing network in chronic cocaine smokers and matched control participants during a cocaine-cue exposure task. Specifically, cocaine-smoking (15M; 5F) and healthy control (13M; 4F) participants viewed cocaine and neutral cues while in the scanner (a Siemens 3 T magnet). We examined whole brain activation, including activation related to drug-cue processing. Time series data extracted from ROIs determined through our General Linear Model (GLM) analysis and prior publications were used as input to IMaGES, a computationally powerful Bayesian search algorithm. During cocaine-cue exposure, cocaine users showed a particular feed-forward effective connectivity pattern between the ROIs of the drug-cue processing network (amygdala → hippocampus → dorsal striatum → insula → medial frontal cortex, dorsolateral prefrontal cortex, anterior cingulate cortex) that was not present when the controls viewed the cocaine cues. Cocaine craving ratings positively correlated with the strength of the causal influence of the insula on the dorsolateral prefrontal cortex in cocaine users. This study is the first demonstration of a causal interaction between ROIs within the drug-cue processing network in cocaine users. This study provides insight into the mechanism underlying continued substance use and has implications for monitoring treatment response.

Brain network response underlying decisions about abstract reinforcers.

Decision making studies typically use tasks that involve concrete action-outcome contingencies, in which subjects do something and get something. No studies have addressed decision making involving abstract reinforcers, where there are no action-outcome contingencies and choices are entirely hypothetical. The present study examines these kinds of choices, as well as whether the same biases that exist for concrete reinforcer decisions, specifically framing effects, also apply during abstract reinforcer decisions. We use both General Linear Model as well as Bayes network connectivity analysis using the Independent Multi-sample Greedy Equivalence Search (IMaGES) algorithm to examine network response underlying choices for abstract reinforcers under positive and negative framing. We find for the first time that abstract reinforcer decisions activate the same network of brain regions as concrete reinforcer decisions, including the striatum, insula, anterior cingulate, and VMPFC, results that are further supported via comparison to a meta-analysis of decision making studies. Positive and negative framing activated different parts of this network, with stronger activation in VMPFC during negative framing and in DLPFC during positive, suggesting different decision making pathways depending on frame. These results were further clarified using connectivity analysis, which revealed stronger connections between anterior cingulate, insula, and accumbens during negative framing compared to positive. Taken together, these results suggest that not only do abstract reinforcer decisions rely on the same brain substrates as concrete reinforcers, but that the response underlying framing effects on abstract reinforcers also resemble those for concrete reinforcers, specifically increased limbic system connectivity during negative frames.

Modeling activation and effective connectivity of VWFA in same script bilinguals.

Previous neuroimaging research revealed a small area in the inferior occipito-temporal cortex (VWFA), which seems to be involved in recognition of written words. The specialized response of the VWFA to words could result from repeated exposure to print in the course of functional fine-tuning of the brain. Research with bilingual speakers holds promise in helping to reveal response properties of the VWFA by assessing its sensitivity to language proficiency, word-form similarity, and meaning overlap across two languages. Using fMRI, we compared VWFA activity for cognate and homograph prime-target pairs in a group of fluent Spanish-English speakers. Cognates share form and meaning in two languages, while homographs only share form. Relative to baseline, the VWFA showed repetition suppression to pairs of homographs, but not to pairs of cognates, suggesting that this area is sensitive to word meaning. The different response to cognates and homographs was only observed when English was the prime language and Spanish was the target language. To help explain this result we compared patterns of effective connectivity between the VWFA and other parts of the reading network implicated in semantic and phonological processing. Our neural models showed that English targets engaged a direct ventral route from the VWFA to the frontal lobe and Spanish targets engaged an indirect dorsal route. Considering that frontal cortex has been implicated in semantic processing, a direct connection to this area could signal a fast and automatic access to meaning and would facilitate early semantic influences in visual word recognition.

Atypical effective connectivity of social brain networks in individuals with autism.

Failing to engage in joint attention is a strong marker of impaired social cognition associated with autism spectrum disorder (ASD). The goal of this study was to localize the source of impaired joint attention in individuals with ASD by examining both behavioral and fMRI data collected during various tasks involving eye gaze, directional cuing, and face processing. The tasks were designed to engage three brain networks associated with social cognition [face processing, theory of mind (TOM), and action understanding]. The behavioral results indicate that even high-functioning individuals with ASD perform less accurately and more slowly than neurotypical (NT) controls when processing eyes, but not when processing a directional cue (an arrow) that did not involve eyes. Behavioral differences between the NT and ASD groups were consistent with differences in the effective connectivity of FACE, TOM, and ACTION networks. An independent multiple-sample greedy equivalence search was used to examine these social brain networks and found that whereas NTs produced stable patterns of response across tasks designed to engage a given brain network, ASD participants did not. Moreover, ASD participants recruited all three networks in a manner highly dissimilar to that of NTs. These results extend a growing literature that describes disruptions in general brain connectivity in individuals with autism by targeting specific networks hypothesized to underlie the social cognitive impairments observed in these individuals.

Dynamic changes in the medial temporal lobe during incidental learning of object-location associations.

The role of the medial temporal lobe (MTL) in associative memory encoding has been the focus of many memory experiments. However, there has been surprisingly little investigation of whether the contributions of different MTL subregions (amygdala, hippocampus [HPC], parahippocampal [PHc], perirhinal cortex [PRc], and temporal polar cortex [TPc]) shift across multiple presentations during associative encoding. We examined this issue using event-related functional magnetic resonance imaging and a multivoxel pattern classification analysis. Subjects performed a visual search task, becoming faster with practice to locate objects whose locations were held constant across trials. The classification analysis implicated right HPC and amygdala early in the task when the speed-up from trial to trial was greatest. The same analysis implicated right PRc and TPc late in learning when speed-up was minimal. These results suggest that associative encoding relies on complex patterns of neural activity in MTL that cannot be expressed by simple increases or decreases of blood oxygenation level-dependent signal during learning. Involvement of MTL subregions during encoding of object-location associations depends on the nature of the learning phase. Right HPC and amygdala support active integration of object and location information, while right PRc and TPc are involved when object and spatial representations become unitized into a single representation.

Multi-subject search correctly identifies causal connections and most causal directions in the DCM models of the Smith et al. simulation study.

Smith et al. report a large study of the accuracy of 38 search procedures for recovering effective connections in simulations of DCM models under 28 different conditions. Their results are disappointing: no method reliably finds and directs connections without large false negatives, large false positives, or both. Using multiple subject inputs, we apply a previously published search algorithm, IMaGES, and novel orientation algorithms, LOFS, in tandem to all of the simulations of DCM models described by Smith et al. (2011). We find that the procedures accurately identify effective connections in almost all of the conditions that Smith et al. simulated and, in most conditions, direct causal connections with precision greater than 90% and recall greater than 80%.

High-resolution imaging of the fusiform face area (FFA) using multivariate non-linear classifiers shows diagnosticity for non-face categories.

Does the "fusiform face area" (FFA) code only for faces? This question continues to elude the neuroimaging field due to at least two kinds of problems: first, the relatively low spatial resolution of fMRI in which the FFA was defined and second, the potential bias inherent in prevailing statistical methods for analyzing the actual diagnosticity of cortical tissue. Using high-resolution (1 mm × 1 mm × 1 mm) imaging data of the fusiform face area (FFA) from 4 subjects who had categorized images as 'animal', 'car', 'face', or 'sculpture', we used multivariate linear and non-linear classifiers to decode the resultant voxel patterns. Prior to identifying the appropriate classifier we performed exploratory analysis to determine the nature of the distributions over classes and the voxel intensity pattern structure between classes. The FFA was visualized using non-metric multidimensional scaling revealing "string-like" sequences of voxels, which appeared in small non-contiguous clusters of categories, intertwined with other categories. Since this analysis suggested that feature space was highly non-linear, we trained various statistical classifiers on the class-conditional distributions (labelled) and separated the four categories with 100% reliability (over replications) and generalized to out-of-sample cases with high significance (up to 50%; p<.000001, chance=25%). The increased noise inherent in high-resolution neuroimaging data relative to standard resolution resisted any further gains in category performance above ~60% (with FACE category often having the highest bias per category) even coupled with various feature extraction/selection methods. A sensitivity/diagnosticity analysis for each classifier per voxel showed: (1) reliable (with S.E.<3%) sensitivity present throughout the FFA for all 4 categories, and (2) showed multi-selectivity; that is, many voxels were selective for more than one category with some high diagnosticity but at submaximal intensity. This work is clearly consistent with the characterization of the FFA as a distributed, object-heterogeneous similarity structure and bolsters the view that the FFA response to "FACE" stimuli in standard resolution may be primarily due to a linear bias, which has resulted from an averaging artefact.

Implicit memory for object locations depends on reactivation of encoding-related brain regions.

This study explored the correspondence between implicit memory and the reactivation of encoding-related brain regions. By using a classification method, we examined whether reactivation reflects only the similarities between study and test or voxels at the reactivated regions are diagnostic of facilitation in the implicit memory task. A simple detection task served as incidental encoding of object-location pairings. A subsequent visual search task served as the indirect (implicit) test of memory. Subjects did not know that their memory would be tested. Half of the subjects were unaware that some stimuli in the search task are the same as those that had appeared during the detection task. Another group of subjects was made aware of this relationship at the onset of the visual search task. Memory performance was superior for the study-test aware, compared to study-test unaware, subjects. Brain reactivation was calculated using a conjunction analysis implemented through overlaying the neural activity at encoding and testing. The conjunction analysis revealed that implicit memory in both groups of subjects was associated with reactivation of parietal and occipital brain regions. We were able to classify study-test aware and study-test unaware subjects based on the per-voxel reactivation values representing the neural dynamics between encoding and test. The classification results indicate that neural dynamics between encoding and test accounts for the differences in implicit memory. Overall, our study demonstrates that implicit memory performance requires and depends upon reactivation of encoding-related brain regions.

Decoding the large-scale structure of brain function by classifying mental States across individuals.

Brain-imaging research has largely focused on localizing patterns of activity related to specific mental processes, but recent work has shown that mental states can be identified from neuroimaging data using statistical classifiers. We investigated whether this approach could be extended to predict the mental state of an individual using a statistical classifier trained on other individuals, and whether the information gained in doing so could provide new insights into how mental processes are organized in the brain. Using a variety of classifier techniques, we achieved cross-validated classification accuracy of 80% across individuals (chance = 13%). Using a neural network classifier, we recovered a low-dimensional representation common to all the cognitive-perceptual tasks in our data set, and we used an ontology of cognitive processes to determine the cognitive concepts most related to each dimension. These results revealed a small organized set of large-scale networks that map cognitive processes across a highly diverse set of mental tasks, suggesting a novel way to characterize the neural basis of cognition.