Human Visual Cortex and Deep Convolutional Neural Network Care Deeply about Object Background.

Deep convolutional neural networks (DCNNs) are able to partially predict brain activity during object categorization tasks, but factors contributing to this predictive power are not fully understood. Our study aimed to investigate the factors contributing to the predictive power of DCNNs in object categorization tasks. We compared the activity of four DCNN architectures with EEG recordings obtained from 62 human participants during an object categorization task. Previous physiological studies on object categorization have highlighted the importance of figure-ground segregation-the ability to distinguish objects from their backgrounds. Therefore, we investigated whether figure-ground segregation could explain the predictive power of DCNNs. Using a stimulus set consisting of identical target objects embedded in different backgrounds, we examined the influence of object background versus object category within both EEG and DCNN activity. Crucially, the recombination of naturalistic objects and experimentally controlled backgrounds creates a challenging and naturalistic task, while retaining experimental control. Our results showed that early EEG activity (< 100 msec) and early DCNN layers represent object background rather than object category. We also found that the ability of DCNNs to predict EEG activity is primarily influenced by how both systems process object backgrounds, rather than object categories. We demonstrated the role of figure-ground segregation as a potential prerequisite for recognition of object features, by contrasting the activations of trained and untrained (i.e., random weights) DCNNs. These findings suggest that both human visual cortex and DCNNs prioritize the segregation of object backgrounds and target objects to perform object categorization. Altogether, our study provides new insights into the mechanisms underlying object categorization as we demonstrated that both human visual cortex and DCNNs care deeply about object background.

Mechanisms of human dynamic object recognition revealed by sequential deep neural networks.

Humans can quickly recognize objects in a dynamically changing world. This ability is showcased by the fact that observers succeed at recognizing objects in rapidly changing image sequences, at up to 13 ms/image. To date, the mechanisms that govern dynamic object recognition remain poorly understood. Here, we developed deep learning models for dynamic recognition and compared different computational mechanisms, contrasting feedforward and recurrent, single-image and sequential processing as well as different forms of adaptation. We found that only models that integrate images sequentially via lateral recurrence mirrored human performance (N = 36) and were predictive of trial-by-trial responses across image durations (13-80 ms/image). Importantly, models with sequential lateral-recurrent integration also captured how human performance changes as a function of image presentation durations, with models processing images for a few time steps capturing human object recognition at shorter presentation durations and models processing images for more time steps capturing human object recognition at longer presentation durations. Furthermore, augmenting such a recurrent model with adaptation markedly improved dynamic recognition performance and accelerated its representational dynamics, thereby predicting human trial-by-trial responses using fewer processing resources. Together, these findings provide new insights into the mechanisms rendering object recognition so fast and effective in a dynamic visual world.

Testing, explaining, and exploring models of facial expressions of emotions.

Models are the hallmark of mature scientific inquiry. In psychology, this maturity has been reached in a pervasive question-what models best represent facial expressions of emotion? Several hypotheses propose different combinations of facial movements [action units (AUs)] as best representing the six basic emotions and four conversational signals across cultures. We developed a new framework to formalize such hypotheses as predictive models, compare their ability to predict human emotion categorizations in Western and East Asian cultures, explain the causal role of individual AUs, and explore updated, culture-accented models that improve performance by reducing a prevalent Western bias. Our predictive models also provide a noise ceiling to inform the explanatory power and limitations of different factors (e.g., AUs and individual differences). Thus, our framework provides a new approach to test models of social signals, explain their predictive power, and explore their optimization, with direct implications for theory development.

A Critical Test of Deep Convolutional Neural Networks' Ability to Capture Recurrent Processing in the Brain Using Visual Masking.

Recurrent processing is a crucial feature in human visual processing supporting perceptual grouping, figure-ground segmentation, and recognition under challenging conditions. There is a clear need to incorporate recurrent processing in deep convolutional neural networks, but the computations underlying recurrent processing remain unclear. In this article, we tested a form of recurrence in deep residual networks (ResNets) to capture recurrent processing signals in the human brain. Although ResNets are feedforward networks, they approximate an excitatory additive form of recurrence. Essentially, this form of recurrence consists of repeating excitatory activations in response to a static stimulus. Here, we used ResNets of varying depths (reflecting varying levels of recurrent processing) to explain EEG activity within a visual masking paradigm. Sixty-two humans and 50 artificial agents (10 ResNet models of depths -4, 6, 10, 18, and 34) completed an object categorization task. We show that deeper networks explained more variance in brain activity compared with shallower networks. Furthermore, all ResNets captured differences in brain activity between unmasked and masked trials, with differences starting at ∼98 msec (from stimulus onset). These early differences indicated that EEG activity reflected "pure" feedforward signals only briefly (up to ∼98 msec). After ∼98 msec, deeper networks showed a significant increase in explained variance, which peaks at ∼200 msec, but only within unmasked trials, not masked trials. In summary, we provided clear evidence that excitatory additive recurrent processing in ResNets captures some of the recurrent processing in humans.

Arousal state affects perceptual decision-making by modulating hierarchical sensory processing in a large-scale visual system model.

Arousal levels strongly affect task performance. Yet, what arousal level is optimal for a task depends on its difficulty. Easy task performance peaks at higher arousal levels, whereas performance on difficult tasks displays an inverted U-shape relationship with arousal, peaking at medium arousal levels, an observation first made by Yerkes and Dodson in 1908. It is commonly proposed that the noradrenergic locus coeruleus system regulates these effects on performance through a widespread release of noradrenaline resulting in changes of cortical gain. This account, however, does not explain why performance decays with high arousal levels only in difficult, but not in simple tasks. Here, we present a mechanistic model that revisits the Yerkes-Dodson effect from a sensory perspective: a deep convolutional neural network augmented with a global gain mechanism reproduced the same interaction between arousal state and task difficulty in its performance. Investigating this model revealed that global gain states differentially modulated sensory information encoding across the processing hierarchy, which explained their differential effects on performance on simple versus difficult tasks. These findings offer a novel hierarchical sensory processing account of how, and why, arousal state affects task performance.

Leveraging Spiking Deep Neural Networks to Understand the Neural Mechanisms Underlying Selective Attention.

Spatial attention enhances sensory processing of goal-relevant information and improves perceptual sensitivity. Yet, the specific neural mechanisms underlying the effects of spatial attention on performance are still contested. Here, we examine different attention mechanisms in spiking deep convolutional neural networks. We directly contrast effects of precision (internal noise suppression) and two different gain modulation mechanisms on performance on a visual search task with complex real-world images. Unlike standard artificial neurons, biological neurons have saturating activation functions, permitting implementation of attentional gain as gain on a neuron's input or on its outgoing connection. We show that modulating the connection is most effective in selectively enhancing information processing by redistributing spiking activity and by introducing additional task-relevant information, as shown by representational similarity analyses. Precision only produced minor attentional effects in performance. Our results, which mirror empirical findings, show that it is possible to adjudicate between attention mechanisms using more biologically realistic models and natural stimuli.

Robust BOLD Responses to Faces But Not to Conditioned Threat: Challenging the Amygdala's Reputation in Human Fear and Extinction Learning.

Most of our knowledge about human emotional memory comes from animal research. Based on this work, the amygdala is often labeled the brain's "fear center", but it is unclear to what degree neural circuitries underlying fear and extinction learning are conserved across species. Neuroimaging studies in humans yield conflicting findings, with many studies failing to show amygdala activation in response to learned threat. Such null findings are often treated as resulting from MRI-specific problems related to measuring deep brain structures. Here we test this assumption in a mega-analysis of three studies on fear acquisition (n = 98; 68 female) and extinction learning (n = 79; 53 female). The conditioning procedure involved the presentation of two pictures of faces and two pictures of houses: one of each pair was followed by an electric shock [a conditioned stimulus (CS+)], the other one was never followed by a shock (CS-), and participants were instructed to learn these contingencies. Results revealed widespread responses to the CS+ compared with the CS- in the fear network, including anterior insula, midcingulate cortex, thalamus, and bed nucleus of the stria terminalis, but not the amygdala, which actually responded stronger to the CS- Results were independent of spatial smoothing, and of individual differences in trait anxiety and conditioned pupil responses. In contrast, robust amygdala activation distinguished faces from houses, refuting the idea that a poor signal could account for the absence of effects. Moving forward, we suggest that, apart from imaging larger samples at higher resolution, alternative statistical approaches may be used to identify cross-species similarities in fear and extinction learning.SIGNIFICANCE STATEMENT The science of emotional memory provides the foundation of numerous theories on psychopathology, including stress and anxiety disorders. This field relies heavily on animal research, which suggests a central role of the amygdala in fear learning and memory. However, this finding is not strongly corroborated by neuroimaging evidence in humans, and null findings are too easily explained away by methodological limitations inherent to imaging deep brain structures. In a large nonclinical sample, we find widespread BOLD activation in response to learned fear, but not in the amygdala. A poor signal could not account for the absence of effects. While these findings do not disprove the involvement of the amygdala in human fear learning, they challenge its typical portrayals and illustrate the complexities of translational science.

Visual features drive the category-specific impairments on categorization tasks in a patient with object agnosia.

Object and scene recognition both require mapping of incoming sensory information to existing conceptual knowledge about the world. A notable finding in brain-damaged patients is that they may show differentially impaired performance for specific categories, such as for "living exemplars". While numerous patients with category-specific impairments have been reported, the explanations for these deficits remain controversial. In the current study, we investigate the ability of a brain injured patient with a well-established category-specific impairment of semantic memory to perform two categorization experiments: 'natural' vs. 'manmade' scenes (experiment 1) and objects (experiment 2). Our findings show that the pattern of categorical impairment does not respect the natural versus manmade distinction. This suggests that the impairments may be better explained by differences in visual features, rather than by category membership. Using Deep Convolutional Neural Networks (DCNNs) as 'artificial animal models' we further explored this idea. Results indicated that DCNNs with 'lesions' in higher order layers showed similar response patterns, with decreased relative performance for manmade scenes (experiment 1) and natural objects (experiment 2), even though they have no semantic category knowledge, apart from a mapping between pictures and labels. Collectively, these results suggest that the direction of category-effects to a large extent depends, at least in MS' case, on the degree of perceptual differentiation called for, and not semantic knowledge.

On the necessity of recurrent processing during object recognition: it depends on the need for scene segmentation.

While feed-forward activity may suffice for recognizing objects in isolation, additional visual operations that aid object recognition might be needed for real-world scenes. One such additional operation is figure-ground segmentation; extracting the relevant features and locations of the target object while ignoring irrelevant features. In this study of 60 human participants (female and male), we show objects on backgrounds of increasing complexity to investigate whether recurrent computations are increasingly important for segmenting objects from more complex backgrounds. Three lines of evidence show that recurrent processing is critical for recognition of objects embedded in complex scenes. First, behavioral results indicated a greater reduction in performance after masking objects presented on more complex backgrounds; with the degree of impairment increasing with increasing background complexity. Second, electroencephalography (EEG) measurements showed clear differences in the evoked response potentials (ERPs) between conditions around time points beyond feed-forward activity and exploratory object decoding analyses based on the EEG signal indicated later decoding onsets for objects embedded in more complex backgrounds. Third, Deep Convolutional Neural Network performance confirmed this interpretation; feed-forward and less deep networks showed a higher degree of impairment in recognition for objects in complex backgrounds compared to recurrent and deeper networks. Together, these results support the notion that recurrent computations drive figure-ground segmentation of objects in complex scenes.SIGNIFICANCE STATEMENTThe incredible speed of object recognition suggests that it relies purely on a fast feed-forward build-up of perceptual activity. However, this view is contradicted by studies showing that disruption of recurrent processing leads to decreased object recognition performance. Here we resolve this issue by showing that how object recognition is resolved, and whether recurrent processing is crucial, depends on the context in which it is presented. For objects presented in isolation or in 'simple' environments, feed-forward activity could be sufficient for successful object recognition. However, when the environment is more complex, additional processing seems necessary to select the elements that belong to the object, and by that segregate them from the background.

Choosing to view morbid information involves reward circuitry.

People often seek out stories, videos or images that detail death, violence or harm. Considering the ubiquity of this behavior, it is surprising that we know very little about the neural circuits involved in choosing negative information. Using fMRI, the present study shows that choosing intensely negative stimuli engages similar brain regions as those that support extrinsic incentives and "regular" curiosity. Participants made choices to view negative and positive images, based on negative (e.g., a soldier kicks a civilian against his head) and positive (e.g., children throw flower petals at a wedding) verbal cues. We hypothesized that the conflicting, but relatively informative act of choosing to view a negative image, resulted in stronger activation of reward circuitry as opposed to the relatively uncomplicated act of choosing to view a positive stimulus. Indeed, as preregistered, we found that choosing negative cues was associated with activation of the striatum, inferior frontal gyrus, anterior insula, and anterior cingulate cortex, both when contrasting against a passive viewing condition, and when contrasting against positive cues. These findings nuance models of decision-making, valuation and curiosity, and are an important starting point when considering the value of seeking out negative content.

Depth in convolutional neural networks solves scene segmentation.

Feed-forward deep convolutional neural networks (DCNNs) are, under specific conditions, matching and even surpassing human performance in object recognition in natural scenes. This performance suggests that the analysis of a loose collection of image features could support the recognition of natural object categories, without dedicated systems to solve specific visual subtasks. Research in humans however suggests that while feedforward activity may suffice for sparse scenes with isolated objects, additional visual operations ('routines') that aid the recognition process (e.g. segmentation or grouping) are needed for more complex scenes. Linking human visual processing to performance of DCNNs with increasing depth, we here explored if, how, and when object information is differentiated from the backgrounds they appear on. To this end, we controlled the information in both objects and backgrounds, as well as the relationship between them by adding noise, manipulating background congruence and systematically occluding parts of the image. Results indicate that with an increase in network depth, there is an increase in the distinction between object- and background information. For more shallow networks, results indicated a benefit of training on segmented objects. Overall, these results indicate that, de facto, scene segmentation can be performed by a network of sufficient depth. We conclude that the human brain could perform scene segmentation in the context of object identification without an explicit mechanism, by selecting or "binding" features that belong to the object and ignoring other features, in a manner similar to a very deep convolutional neural network.

Low-level image statistics in natural scenes influence perceptual decision-making.

A fundamental component of interacting with our environment is gathering and interpretation of sensory information. When investigating how perceptual information influences decision-making, most researchers have relied on manipulated or unnatural information as perceptual input, resulting in findings that may not generalize to real-world scenes. Unlike simplified, artificial stimuli, real-world scenes contain low-level regularities that are informative about the structural complexity, which the brain could exploit. In this study, participants performed an animal detection task on low, medium or high complexity scenes as determined by two biologically plausible natural scene statistics, contrast energy (CE) or spatial coherence (SC). In experiment 1, stimuli were sampled such that CE and SC both influenced scene complexity. Diffusion modelling showed that the speed of information processing was affected by low-level scene complexity. Experiment 2a/b refined these observations by showing how isolated manipulation of SC resulted in weaker but comparable effects, with an additional change in response boundary, whereas manipulation of only CE had no effect. Overall, performance was best for scenes with intermediate complexity. Our systematic definition quantifies how natural scene complexity interacts with decision-making. We speculate that CE and SC serve as an indication to adjust perceptual decision-making based on the complexity of the input.

Neurocognitive Underpinnings of Aggressive Predation in Economic Contests.

Competitions are part and parcel of daily life and require people to invest time and energy to gain advantage over others and to avoid (the risk of) falling behind. Whereas the behavioral mechanisms underlying competition are well documented, its neurocognitive underpinnings remain poorly understood. We addressed this using neuroimaging and computational modeling of individual investment decisions aimed at exploiting one's counterpart ("attack") or at protecting against exploitation by one's counterpart ("defense"). Analyses revealed that during attack relative to defense (i) individuals invest less and are less successful; (ii) computations of expected reward are strategically more sophisticated (reasoning level k = 4 vs. k = 3 during defense); (iii) ventral striatum activity tracks reward prediction errors; (iv) risk prediction errors were not correlated with neural activity in either ROI or whole-brain analyses; and (v) successful exploitation correlated with neural activity in the bilateral ventral striatum, left OFC, left anterior insula, left TPJ, and lateral occipital cortex. We conclude that, in economic contests, coming out ahead (vs. not falling behind) involves sophisticated strategic reasoning that engages both reward and value computation areas and areas associated with theory of mind.

Exome Chip Meta-analysis Fine Maps Causal Variants and Elucidates the Genetic Architecture of Rare Coding Variants in Smoking and Alcohol Use.

BACKGROUND: Smoking and alcohol use have been associated with common genetic variants in multiple loci. Rare variants within these loci hold promise in the identification of biological mechanisms in substance use. Exome arrays and genotype imputation can now efficiently genotype rare nonsynonymous and loss of function variants. Such variants are expected to have deleterious functional consequences and to contribute to disease risk. METHODS: We analyzed ∼250,000 rare variants from 16 independent studies genotyped with exome arrays and augmented this dataset with imputed data from the UK Biobank. Associations were tested for five phenotypes: cigarettes per day, pack-years, smoking initiation, age of smoking initiation, and alcoholic drinks per week. We conducted stratified heritability analyses, single-variant tests, and gene-based burden tests of nonsynonymous/loss-of-function coding variants. We performed a novel fine-mapping analysis to winnow the number of putative causal variants within associated loci. RESULTS: Meta-analytic sample sizes ranged from 152,348 to 433,216, depending on the phenotype. Rare coding variation explained 1.1% to 2.2% of phenotypic variance, reflecting 11% to 18% of the total single nucleotide polymorphism heritability of these phenotypes. We identified 171 genome-wide associated loci across all phenotypes. Fine mapping identified putative causal variants with double base-pair resolution at 24 of these loci, and between three and 10 variants for 65 loci. Twenty loci contained rare coding variants in the 95% credible intervals. CONCLUSIONS: Rare coding variation significantly contributes to the heritability of smoking and alcohol use. Fine-mapping genome-wide association study loci identifies specific variants contributing to the biological etiology of substance use behavior.

How to control for confounds in decoding analyses of neuroimaging data.

Over the past decade, multivariate "decoding analyses" have become a popular alternative to traditional mass-univariate analyses in neuroimaging research. However, a fundamental limitation of using decoding analyses is that it remains ambiguous which source of information drives decoding performance, which becomes problematic when the to-be-decoded variable is confounded by variables that are not of primary interest. In this study, we use a comprehensive set of simulations as well as analyses of empirical data to evaluate two methods that were previously proposed and used to control for confounding variables in decoding analyses: post hoc counterbalancing and confound regression. In our empirical analyses, we attempt to decode gender from structural MRI data while controlling for the confound "brain size". We show that both methods introduce strong biases in decoding performance: post hoc counterbalancing leads to better performance than expected (i.e., positive bias), which we show in our simulations is due to the subsampling process that tends to remove samples that are hard to classify or would be wrongly classified; confound regression, on the other hand, leads to worse performance than expected (i.e., negative bias), even resulting in significant below chance performance in some realistic scenarios. In our simulations, we show that below chance accuracy can be predicted by the variance of the distribution of correlations between the features and the target. Importantly, we show that this negative bias disappears in both the empirical analyses and simulations when the confound regression procedure is performed in every fold of the cross-validation routine, yielding plausible (above chance) model performance. We conclude that, from the various methods tested, cross-validated confound regression is the only method that appears to appropriately control for confounds which thus can be used to gain more insight into the exact source(s) of information driving one's decoding analysis.

The influence of acoustic startle probes on fear learning in humans.

Even though human fear-conditioning involves affective learning as well as expectancy learning, most studies assess only one of the two distinct processes. Commonly used read-outs of associative fear learning are the fear-potentiated startle reflex (FPS), pupil dilation and US-expectancy ratings. FPS is thought to reflect the affective aspect of fear learning, while pupil dilation reflects a general arousal response. However, in order to measure FPS, aversively loud acoustic probes are presented during conditioning, which might in itself exert an effect on fear learning. Here we tested the effect of startle probes on fear learning by comparing brain activation (fMRI), pupil dilation and US-expectancy ratings with and without acoustic startle probes within subjects. Regardless of startle probes, fear conditioning resulted in enhanced dACC, insula and ventral striatum activation. Interaction analyses showed that startle probes diminished differential pupil dilation between CS+ and CS- due to increased pupil responses to CS-. A trend significant interaction effect was observed for US-expectancy and amygdala activation. Startle probes affect differential fear learning by impeding safety learning, as measured with pupil dilation, a read-out of the cognitive component of fear learning. However, we observed no significant effect of acoustic startle probes on other measures of fear learning.

Pupil mimicry promotes trust through the theory-of-mind network.

The human eye can provide powerful insights into the emotions and intentions of others; however, how pupillary changes influence observers' behavior remains largely unknown. The present fMRI-pupillometry study revealed that when the pupils of interacting partners synchronously dilate, trust is promoted, which suggests that pupil mimicry affiliates people. Here we provide evidence that pupil mimicry modulates trust decisions through the activation of the theory-of-mind network (precuneus, temporo-parietal junction, superior temporal sulcus, and medial prefrontal cortex). This network was recruited during pupil-dilation mimicry compared with interactions without mimicry or compared with pupil-constriction mimicry. Furthermore, the level of theory-of-mind engagement was proportional to individual's susceptibility to pupil-dilation mimicry. These data reveal a fundamental mechanism by which an individual's pupils trigger neurophysiological responses within an observer: when interacting partners synchronously dilate their pupils, humans come to feel reflections of the inner states of others, which fosters trust formation.

Scene complexity modulates degree of feedback activity during object detection in natural scenes.

Selective brain responses to objects arise within a few hundreds of milliseconds of neural processing, suggesting that visual object recognition is mediated by rapid feed-forward activations. Yet disruption of neural responses in early visual cortex beyond feed-forward processing stages affects object recognition performance. Here, we unite these discrepant findings by reporting that object recognition involves enhanced feedback activity (recurrent processing within early visual cortex) when target objects are embedded in natural scenes that are characterized by high complexity. Human participants performed an animal target detection task on natural scenes with low, medium or high complexity as determined by a computational model of low-level contrast statistics. Three converging lines of evidence indicate that feedback was selectively enhanced for high complexity scenes. First, functional magnetic resonance imaging (fMRI) activity in early visual cortex (V1) was enhanced for target objects in scenes with high, but not low or medium complexity. Second, event-related potentials (ERPs) evoked by target objects were selectively enhanced at feedback stages of visual processing (from ~220 ms onwards) for high complexity scenes only. Third, behavioral performance for high complexity scenes deteriorated when participants were pressed for time and thus less able to incorporate the feedback activity. Modeling of the reaction time distributions using drift diffusion revealed that object information accumulated more slowly for high complexity scenes, with evidence accumulation being coupled to trial-to-trial variation in the EEG feedback response. Together, these results suggest that while feed-forward activity may suffice to recognize isolated objects, the brain employs recurrent processing more adaptively in naturalistic settings, using minimal feedback for simple scenes and increasing feedback for complex scenes.

Sparse Computation in Adaptive Spiking Neural Networks.

Artificial Neural Networks (ANNs) are bio-inspired models of neural computation that have proven highly effective. Still, ANNs lack a natural notion of time, and neural units in ANNs exchange analog values in a frame-based manner, a computationally and energetically inefficient form of communication. This contrasts sharply with biological neurons that communicate sparingly and efficiently using isomorphic binary spikes. While Spiking Neural Networks (SNNs) can be constructed by replacing the units of an ANN with spiking neurons (Cao et al., 2015; Diehl et al., 2015) to obtain reasonable performance, these SNNs use Poisson spiking mechanisms with exceedingly high firing rates compared to their biological counterparts. Here we show how spiking neurons that employ a form of neural coding can be used to construct SNNs that match high-performance ANNs and match or exceed state-of-the-art in SNNs on important benchmarks, while requiring firing rates compatible with biological findings. For this, we use spike-based coding based on the firing rate limiting adaptation phenomenon observed in biological spiking neurons. This phenomenon can be captured in fast adapting spiking neuron models, for which we derive the effective transfer function. Neural units in ANNs trained with this transfer function can be substituted directly with adaptive spiking neurons, and the resulting Adaptive SNNs (AdSNNs) can carry out competitive classification in deep neural networks without further modifications. Adaptive spike-based coding additionally allows for the dynamic control of neural coding precision: we show empirically how a simple model of arousal in AdSNNs further halves the average required firing rate and this notion naturally extends to other forms of attention as studied in neuroscience. AdSNNs thus hold promise as a novel and sparsely active model for neural computation that naturally fits to temporally continuous and asynchronous applications.