Higher predicted brain age difference has been associated with several psychiatric disorders. Generalized anxiety disorder (GAD) is associated with markers of accelerated aging. In this study, we determined brain predicted age difference (PAD) in individuals with GAD and healthy controls (HC) as well as group differences in PAD variability using voxel-wise structural MRI. The training dataset included 3,511 controls, and the testing dataset included 1,595 individuals with GAD and 4,552 HC from the ENIGMA-Anxiety GAD Working Group. A convolutional neural network model using four input modalities per subject and a model ensemble approach was used to predict brain age. The PAD was then calculated by subtracting chronological age. Model performance was consistent with other image-based brain age prediction models with similar accuracy across the training set (mean absolute error (MAE) = 2.95 years) and HC in the testing set (MAE = 2.94). We found no evidence of accelerated brain aging in individuals with GAD, though we did find evidence for greater variation in PAD for individuals with GAD (Levene's test: W = 442.98, p < .001) and evidence for greater variability in PAD of those with GAD over 25 years of age. No relationships between PAD and clinical or demographic measures were found. To conclude, using large training and testing samples, the study found no significant association between GAD and PAD, although individuals with GAD had greater heterogeneity in brain-predicted age.

Federated learning (FL) is an approach to training machine learning models that takes advantage of multiple distributed datasets while maintaining data privacy and reducing communication costs associated with sharing local datasets. Aggregation strategies have been developed to pool or fuse the weights and biases of distributed deterministic models; however, modern deterministic deep learning (DL) models are often poorly calibrated and lack the ability to communicate a measure of epistemic uncertainty in prediction, which is desirable for remote sensing platforms and safety-critical applications. Conversely, Bayesian DL models are often well calibrated and capable of quantifying and communicating a measure of epistemic uncertainty along with a competitive prediction accuracy. Unfortunately, because the weights and biases in Bayesian DL models are defined by a probability distribution, simple application of the aggregation methods associated with FL schemes for deterministic models is either impossible or results in sub-optimal performance. In this work, we use independent and identically distributed (IID) and non-IID partitions of the CIFAR-10 dataset and a fully variational ResNet-20 architecture to analyze six different aggregation strategies for Bayesian DL models. Additionally, we analyze the traditional federated averaging approach applied to an approximate Bayesian Monte Carlo dropout model as a lightweight alternative to more complex variational inference methods in FL. We show that aggregation strategy is a key hyperparameter in the design of a Bayesian FL system with downstream effects on accuracy, calibration, uncertainty quantification, training stability, and client compute requirements.

One of the central objectives of contemporary neuroimaging research is to create predictive models that can disentangle the connection between patterns of functional connectivity across the entire brain and various behavioral traits. Previous studies have shown that models trained to predict behavioral features from the individual's functional connectivity have modest to poor performance. In this study, we trained models that predict observable individual traits (phenotypes) and their corresponding singular value decomposition (SVD) representations - herein referred to as latent phenotypes from resting state functional connectivity. For this task, we predicted phenotypes in two large neuroimaging datasets: the Human Connectome Project (HCP) and the Philadelphia Neurodevelopmental Cohort (PNC). We illustrate the importance of regressing out confounds, which could significantly influence phenotype prediction. Our findings reveal that both phenotypes and their corresponding latent phenotypes yield similar predictive performance. Interestingly, only the first five latent phenotypes were reliably identified, and using just these reliable phenotypes for predicting phenotypes yielded a similar performance to using all latent phenotypes. This suggests that the predictable information is present in the first latent phenotypes, allowing the remainder to be filtered out without any harm in performance. This study sheds light on the intricate relationship between functional connectivity and the predictability and reliability of phenotypic information, with potential implications for enhancing predictive modeling in the realm of neuroimaging research.

Evasion attacks in adversarial machine learning (AML) have been well-established for multi-class classification problems. However, for multi-label audio classification, several considerations and unique challenges exist in evasion AML that the attacker needs to consider. First, the attacker must cater the attack from the multi-class scenario assumed by the vast majority of evasion attacks to multi-label AML. Second, the attacker should consider creating imper-ceptible adversarial examples to reduce suspicion of the human element by factoring in properties inherent to the audio data set and model. Lastly, with the proliferation of AML techniques, the attacker should assume that the target model will implement a deep neural network (DNN) defense and should attempt a strategy to overcome those defenses. In this paper, we show how an attacker would cater and design an attack against multi-label audio classification using stochastic DNNs, highlighting some considerations for linking AML techniques to acoustic-based metrics.

Deep neural networks (DNNs) are being increasingly used to make predictions from functional magnetic resonance imaging (fMRI) data. However, they are widely seen as uninterpretable “black boxes,” as it can be difficult to discover what input information is used by the DNN in the process, something important in both cognitive neuroscience and clinical applications. A saliency map is a common approach for producing interpretable visualizations of the relative importance of input features for a prediction. However, methods for creating maps often fail due to DNNs being sensitive to input noise or by focusing too much on the input and too little on the model. It is also challenging to evaluate how well saliency maps correspond to the truly relevant input information, as ground truth is not always available. In this paper, we review a variety of methods for producing gradient-based saliency maps and present an adversarial training method we developed to make DNNs robust to input noise, with the goal of improving the quality of DNN saliency maps. We introduce two quantitative evaluation procedures for saliency map methods in fMRI, applicable whenever a DNN or linear model is being trained to decode some information from imaging data. We evaluate the procedures using synthetic data sets, where the complex activation structure is known, and on fMRI data from the Human Connectome Project using saliency maps produced for linear models and DNNs doing task decoding in both settings. Our key finding is that saliency maps produced with different methods vary widely in quality in both synthetic and Human Connectome Project fMRI data, even when those methods have similar prediction performance. We found that training both linear and non-linear models using adversarial noise increased the quality of their saliency maps. We also found that, while some linear models generate good saliency maps in the highly controlled, synthetic data, non-linear DNN methods generate better saliency maps in real-world fMRI data. Finally, our experiments give evidence that combining adversarial training with a complex, non-linear model can improve saliency map quality when compared to several methods commonly used in the literature.

Many stakeholders struggle to make reliances on ML-driven systems due to the risk of harm these systems may cause. Concerns of trustworthiness, unintended social harms, and unacceptable social and ethical violations undermine the promise of ML advancements. Moreover, such risks in complex ML-driven systems present a special challenge as they are often difficult to foresee, arising over periods of time, across populations, and at scale. These risks often arise not from poor ML development decisions or low performance directly but rather emerge through the interactions amongst ML development choices, the context of model use, environmental factors, and the effects of a model on its target. Systems safety engineering is an established discipline with a proven track record of identifying and managing risks even in high-complexity sociotechnical systems. In this work, we apply a state-of-the-art systems safety approach to concrete applications of ML with notable social and ethical risks to demonstrate a systematic means for meeting the assurance requirements needed to argue for safe and trustworthy ML in sociotechnical systems.

A central goal in the cognitive sciences is the development of numerical models for mental representations of object concepts. This paper introduces Variational Interpretable Concept Embeddings (VICE), an approximate Bayesian method for embedding object concepts in a vector space using data collected from humans in a triplet odd-one-out task. VICE uses variational inference to obtain sparse, non-negative representations of object concepts with uncertainty estimates for the embedding values. These estimates are used to automatically select the dimensions that best explain the data. We derive a PAC learning bound for VICE that can be used to estimate generalization performance or determine a sufficient sample size for experimental design. VICE rivals or outperforms its predecessor, SPoSE, at predicting human behavior in the triplet odd-one-out task. Furthermore, VICE's object representations are more reproducible and consistent across random initializations, highlighting the unique advantage of using VICE for deriving interpretable embeddings from human behavior.

In this paper we introduce Variational Inference for Concept Embeddings (VICE), a novel method for learning object concept embeddings from human behavior in an odd-one-out task. We use variational inference to obtain a sparse, non-negative solution, with uncertainty information about each embedding value. We leverage this information in a statistical procedure for selecting the dimensionality of the model, based on hypothesis-testing over a validation set. VICE performs as well or better than previous methods on a variety of criteria: accuracy of predicting human behavior in an odd-one-out task, calibration to (empirical) human choice probabilities, reproducibility of object representations across different random initializations, and superior performance on small datasets. The latter is particularly important in cognitive science, where data collection is expensive. Finally, VICE yields highly interpretable object representations, allowing humans to describe the characteristics being represented by each latent dimension.

Eye tracking provides insights into social processing deficits in autism spectrum disorder (ASD), especially in conjunction with dynamic, naturalistic free-viewing stimuli. However, the question remains whether gaze characteristics, such as preference for specific facial features, can be considered a stable individual trait, particularly in those with ASD. If so, how much data are needed for consistent estimations? To address these questions, we assessed the stability and robustness of gaze preference for facial features as incremental amounts of movie data were introduced for analysis. We trained an artificial neural network to create an object-based segmentation of naturalistic movie clips (14 s each, 7410 frames total). Thirty-three high-functioning individuals with ASD and 36 age- and IQ-equated typically developing individuals (age range: 12-30 years) viewed 22 Hollywood movie clips, each depicting a social interaction. As we evaluated combinations of one, three, five, eight, and 11 movie clips, gaze dwell times on core facial features became increasingly stable at within-subject, within-group, and between-group levels. Using a number of movie clips deemed sufficient by our analysis, we found that individuals with ASD displayed significantly less face-centered gaze (centralized on the nose; p < 0.001) but did not significantly differ from typically developing participants in eye or mouth looking times. Our findings validate gaze preference for specific facial features as a stable individual trait and highlight the possibility of misinterpretation with insufficient data. Additionally, we propose the use of a machine learning approach to stimuli segmentation to quickly and flexibly prepare dynamic stimuli for analysis. Lay Summary Using a data-driven approach to segmenting movie stimuli, we examined varying amounts of data to assess the stability of social gaze in individuals with autism spectrum disorder (ASD). We found a reduction in social fixations in participants with ASD, driven by decreased attention to the center of the face. Our findings further support the validity of gaze preference for face features as a stable individual trait when sufficient data are used.

Deep Reinforcement Learning (Deep RL) agents have in recent years emerged as successful models of animal behavior in a variety of complex learning tasks, as exemplified by Song et al. [2017]. As agents are typically trained to mimic an animal subject, the emphasis in past studies on behavior as a means of evaluating the fitness of models to experimental data is only natural. But the true power of Deep RL agents lies in their ability to learn neural computations and codes that generate a particular behavior—factors that are also of great relevance and interest to computational neuroscience. On that basis, we believe that model evaluation should include an examination of neural representations and validation against neural recordings from animal subjects. In this paper, we introduce a procedure to test hypotheses about the relationship between internal representations of Deep RL agents and those in animal neural recordings. Taking a sequential learning task as a running example, we apply our method and show that the geometry of representations learnt by artificial agents is similar to that of the biological subjects’, and that such similarities are driven by shared information in some latent space. Our method is applicable to any Deep RL agent that learns a Markov Decision Process, and as such enables researchers to assess the suitability of more advanced Deep Learning modules, or map hierarchies of representations to different parts of a circuit in the brain, and help shed light on their function. To demonstrate that point, we conduct an ablation study to deduce that, in the sequential task under consideration, temporal information plays a key role in molding a correct representation of the task.

This short article describes a deep neural network trained to perform automatic segmentation of human body parts in natural scenes. More specifically, we trained a Bayesian SegNet with concrete dropout on the Pascal-Parts dataset to predict whether each pixel in a given frame was part of a person's hair, head, ear, eyebrows, legs, arms, mouth, neck, nose, or torso.

Eye tracking provides insights into social processing and its deficits in disorders such as autism spectrum disorder (ASD), especially in conjunction with dynamic, naturalistic stimuli. However, reliance on manual stimuli segmentation severely limits scalability. We assessed how the amount of available data impacts individual reliability of fixation preference for different facial features, and the effect of this reliability on between-group differences. We trained an artificial neural network to segment 22 Hollywood movie clips (7410 frames). We then analyzed fixation preferences in typically developing participants and participants with ASD as we incrementally introduced movie data for analysis. Although fixations were initially variable, results stabilized as more data was added. Additionally, while those with ASD displayed significantly fewer face-centered fixations ( p <.001), they did not differ in eye or mouth fixations. Our results highlight the validity of treating fixation preferences as a stable individual trait, and the risk of misinterpretation with insufficient data. ### Competing Interest Statement The authors have declared no competing interest.

In this paper, we describe a Bayesian deep neural network (DNN) for predicting FreeSurfer segmentations of structural MRI volumes, in minutes rather than hours. The network was trained and evaluated on a large dataset (n = 11,480), obtained by combining data from more than a hundred different sites, and also evaluated on another completely held-out dataset (n = 418). The network was trained using a novel spike-and-slab dropout-based variational inference approach. We show that, on these datasets, the proposed Bayesian DNN outperforms previously proposed methods, in terms of the similarity between the segmentation predictions and the FreeSurfer labels, and the usefulness of the estimate uncertainty of these predictions. In particular, we demonstrated that the prediction uncertainty of this network at each voxel is a good indicator of whether the network has made an error and that the uncertainty across the whole brain can predict the manual quality control ratings of a scan. The proposed Bayesian DNN method should be applicable to any new network architecture for addressing the segmentation problem.

Collecting the large datasets needed to train deep neural networks can be very difficult, particularly for the many applications for which sharing and pooling data is complicated by practical, ethical, or legal concerns. However, it may be the case that derivative datasets or predictive models developed within individual sites can be shared and combined with fewer restrictions. Training on distributed data and combining the resulting networks is often viewed as continual learning, but these methods require networks to be trained sequentially. In this paper, we introduce distributed weight consolidation (DWC), a continual learning method to consolidate the weights of separate neural networks, each trained on an independent dataset. We evaluated DWC with a brain segmentation case study, where we consolidated dilated convolutional neural networks trained on independent structural magnetic resonance imaging (sMRI) datasets from different sites. We found that DWC led to increased performance on test sets from the different sites, while maintaining generalization performance for a very large and completely independent multi-site dataset, compared to an ensemble baseline.

[This corrects the article DOI: 10.3389/fpsyg.2017.01551.].

The goal of computational neuroscience is to find mechanistic explanations of how the nervous system processes information to give rise to cognitive function and behaviour. At the heart of the field are its models, i.e. mathematical and computational descriptions of the system being studied, which map sensory stimuli to neural responses and/or neural to behavioural responses. These models range from simple to complex. Recently, deep neural networks (DNNs) have come to dominate several domains of artificial intelligence (AI). As the term “neural network” suggests, these models are inspired by biological brains. However, current DNNs neglect many details of biological neural networks. These simplifications contribute to their computational efficiency, enabling them to perform complex feats of intelligence, ranging from perceptual (e.g. visual object and auditory speech recognition) to cognitive tasks (e.g. machine translation), and on to motor control (e.g. playing computer games or controlling a robot arm). In addition to their ability to model complex intelligent behaviours, DNNs excel at predicting neural responses to novel sensory stimuli with accuracies well beyond any other currently available model type. DNNs can have millions of parameters, which are required to capture the domain knowledge needed for successful task performance. Contrary to the intuition that this renders them into impenetrable black boxes, the computational properties of the network units are the result of four directly manipulable elements: input statistics, network structure, functional objective, and learning algorithm. With full access to the activity and connectivity of all units, advanced visualization techniques, and analytic tools to map network representations to neural data, DNNs represent a powerful framework for building task-performing models and will drive substantial insights in computational neuroscience.

The "finger of God" allusion in Luke 11: 20 is usually recognized to draw from explicit OT occurrences of the phrase, particularly those in Exod 8: 15 MT and Deut 9: 10. While not denying the influence of those passages, this article aims to explore and elucidate the unique, but oft-neglected contribution of Dan 5 to Luke 11 by way of "finger of God" imagery.

As deep neural networks (DNNs) are applied to increasingly challenging problems, they will need to be able to represent their own uncertainty. Modelling uncertainty is one of the key features of Bayesian methods. Bayesian DNNs that use dropout-based variational distributions and scale to complex tasks have recently been proposed. We evaluate Bayesian DNNs trained with Bernoulli or Gaussian multiplicative masking of either the units (dropout) or the weights (dropconnect). We compare these Bayesian DNNs ability to represent their uncertainty about their outputs through sampling during inference. We tested the calibration of these Bayesian fully connected and convolutional DNNs on two visual inference tasks (MNIST and CIFAR-10). By adding different levels of Gaussian noise to the test images in z-score space, we assessed how these DNNs represented their uncertainty about regions of input space not covered by the training set. These Bayesian DNNs represented their own uncertainty more accurately than traditional DNNs with a softmax output. We find that sampling of weights, whether Gaussian or Bernoulli, led to more accurate representation of uncertainty compared to sampling of units. However, sampling units using either Gaussian or Bernoulli dropout led to increased convolutional neural network (CNN) classification accuracy. Based on these findings we use both Bernoulli dropout and Gaussian dropconnect concurrently, which approximates the use of a spike-andslab variational distribution. We find that networks with spike-and-slab sampling combine the advantages of the other methods: they classify with high accuracy and robustly represent the uncertainty of their classifications for all tested architectures.