How biological attention mechanisms improve task performance in a large-scale visual system model

Emergence of Emotion Selectivity in Deep Neural Networks Trained to Recognize Visual Objects

Recent neuroimaging studies have shown that the visual cortex plays an important role in representing the affective significance of visual input. The origin of these affect-specific visual representations is debated: they are intrinsic to the visual system versus they arise through reentry from frontal emotion processing structures such as the amygdala. We examined this problem by combining convolutional neural network (CNN) models of the human ventral visual cortex pre-trained on ImageNet with two datasets of affective images. Our results show that (1) in all layers of the CNN models, there were artificial neurons that responded consistently and selectively to neutral, pleasant, or unpleasant images and (2) lesioning these neurons by setting their output to 0 or enhancing these neurons by increasing their gain led to decreased or increased emotion recognition performance respectively. These results support the idea that the visual system may have the intrinsic ability to represent the affective significance of visual input and suggest that CNNs offer a fruitful platform for testing neuroscientific theories.

Emergence of Emotion Selectivity in Deep Neural Networks Trained to Recognize Visual Objects

Recent neuroimaging studies have shown that the visual cortex plays an important role in representing the affective significance of visual input. The origin of these affect-specific visual representations is debated: they are intrinsic to the visual system versus they arise through reentry from frontal emotion processing structures such as the amygdala. We examined this problem by combining convolutional neural network (CNN) models of the human ventral visual cortex pre-trained on ImageNet with two datasets of affective images. Our results show that in all layers of the CNN models, there were artificial neurons that responded consistently and selectively to neutral, pleasant, or unpleasant images and lesioning these neurons by setting their output to zero or enhancing these neurons by increasing their gain led to decreased or increased emotion recognition performance respectively. These results support the idea that the visual system may have the intrinsic ability to represent the affective significance of visual input and suggest that CNNs offer a fruitful platform for testing neuroscientific theories.

Emergent human-like covert attention in feedforward convolutional neural networks

Covert attention allows the selection of locations or features of the visual scene without moving the eyes. Cues and contexts predictive of a target's location orient covert attention and improve perceptual performance. The performance benefits are widely attributed to theories of covert attention as a limited resource, zoom, spotlight, or weighting of visual information. However, such concepts are difficult to map to neuronal populations. We show that a feedforward convolutional neural network (CNN) trained on images to optimize target detection accuracy and with no explicit incorporation of an attention mechanism, a limited resource, or feedback connections learns to utilize cues and contexts in the three most prominent covert attention tasks (Posner cueing, set size effects in search, and contextual cueing) and predicts the cue/context influences on human accuracy. The CNN's cueing/context effects generalize across network training schemes, to peripheral and central pre-cues, discrimination tasks, and reaction time measures, and critically do not vary with reductions in network resources (size). The CNN shows comparable cueing/context effects to a model that optimally uses image information to make decisions (Bayesian ideal observer) but generalizes these effects to cue instances unseen during training. Together, the findings suggest that human-like behavioral signatures of covert attention in the three landmark paradigms might be an emergent property of task accuracy optimization in neuronal populations without positing limited attentional resources. The findings might explain recent behavioral results showing cueing and context effects across a variety of simple organisms with no neocortex, from archerfish to fruit flies.

Grounding neuroscience in behavioral changes using artificial neural networks

Connecting neural activity to function is a common aim in neuroscience. How to define and conceptualize function, however, can vary. Here I focus on grounding this goal in the specific question of how a given change in behavior is produced by a change in neural circuits or activity. Artificial neural network models offer a particularly fruitful format for tackling such questions because they use neural mechanisms to perform complex transformations and produce appropriate behavior. Therefore, they can be a means of causally testing the extent to which a neural change can be responsible for an experimentally observed behavioral change. Furthermore, because the field of interpretability in artificial intelligence has similar aims, neuroscientists can look to interpretability methods for new ways of identifying neural features that drive performance and behaviors.

Targeted V1 comodulation supports task-adaptive sensory decisions

Sensory-guided behavior requires reliable encoding of stimulus information in neural populations, and flexible, task-specific readout. The former has been studied extensively, but the latter remains poorly understood. We introduce a theory for adaptive sensory processing based on functionally-targeted stochastic modulation. We show that responses of neurons in area V1 of monkeys performing a visual discrimination task exhibit low-dimensional, rapidly fluctuating gain modulation, which is stronger in task-informative neurons and can be used to decode from neural activity after few training trials, consistent with observed behavior. In a simulated hierarchical neural network model, such labels are learned quickly and can be used to adapt downstream readout, even after several intervening processing stages. Consistently, we find the modulatory signal estimated in V1 is also present in the activity of simultaneously recorded MT units, and is again strongest in task-informative neurons. These results support the idea that co-modulation facilitates task-adaptive hierarchical information routing.

Improved prediction of behavioral and neural similarity spaces using pruned DNNs

Deep Neural Networks (DNNs) have become an important tool for modeling brain and behavior. One key area of interest has been to apply these networks to model human similarity judgements. Several previous works have used the embeddings from the penultimate layer of vision DNNs and showed that a reweighting of these features improves the fit between human similarity judgments and DNNs. These studies underline the idea that these embeddings form a good basis set but lack the correct level of salience. Here we re-examined the grounds for this idea and on the contrary, we hypothesized that these embeddings, beyond forming a good basis set, also have the correct level of salience to account for similarity judgments. It is just that the huge dimensional embedding needs to be pruned to select those features relevant for the considered domain for which a similarity space is modeled. In Study 1 we supervised DNN pruning based on a subset of human similarity judgments. We found that pruning: i) improved out-of-sample prediction of human similarity judgments from DNN embeddings, ii) produced better alignment with WordNet hierarchy, and iii) retained much higher classification accuracy than reweighting. Study 2 showed that pruning by neurobiological data is highly effective in improving out-of-sample prediction of brain-derived representational dissimilarity matrices from DNN embeddings, at times fleshing out isomorphisms not otherwise observable. Using pruned DNNs, image-level heatmaps can be produced to identify image sections whose features load on dimensions coded by a brain area. Pruning supervised by human brain/behavior therefore effectively identifies alignable dimensions of knowledge between DNNs and humans and constitutes an effective method for understanding the organization of knowledge in neural networks.

The neuroconnectionist research programme

Artificial neural networks (ANNs) inspired by biology are beginning to be widely used to model behavioural and neural data, an approach we call 'neuroconnectionism'. ANNs have been not only lauded as the current best models of information processing in the brain but also criticized for failing to account for basic cognitive functions. In this Perspective article, we propose that arguing about the successes and failures of a restricted set of current ANNs is the wrong approach to assess the promise of neuroconnectionism for brain science. Instead, we take inspiration from the philosophy of science, and in particular from Lakatos, who showed that the core of a scientific research programme is often not directly falsifiable but should be assessed by its capacity to generate novel insights. Following this view, we present neuroconnectionism as a general research programme centred around ANNs as a computational language for expressing falsifiable theories about brain computation. We describe the core of the programme, the underlying computational framework and its tools for testing specific neuroscientific hypotheses and deriving novel understanding. Taking a longitudinal view, we review past and present neuroconnectionist projects and their responses to challenges and argue that the research programme is highly progressive, generating new and otherwise unreachable insights into the workings of the brain.

A rubric for human-like agents and NeuroAI

Researchers across cognitive, neuro- and computer sciences increasingly reference 'human-like' artificial intelligence and 'neuroAI'. However, the scope and use of the terms are often inconsistent. Contributed research ranges widely from mimicking behaviour, to testing machine learning methods as neurally plausible hypotheses at the cellular or functional levels, or solving engineering problems. However, it cannot be assumed nor expected that progress on one of these three goals will automatically translate to progress in others. Here, a simple rubric is proposed to clarify the scope of individual contributions, grounded in their commitments to human-like behaviour, neural plausibility or benchmark/engineering/computer science goals. This is clarified using examples of weak and strong neuroAI and human-like agents, and discussing the generative, corroborate and corrective ways in which the three dimensions interact with one another. The author maintains that future progress in artificial intelligence will need strong interactions across the disciplines, with iterative feedback loops and meticulous validity tests-leading to both known and yet-unknown advances that may span decades to come. This article is part of a discussion meeting issue 'New approaches to 3D vision'.

Efficient coding theory of dynamic attentional modulation

Activity of sensory neurons is driven not only by external stimuli but also by feedback signals from higher brain areas. Attention is one particularly important internal signal whose presumed role is to modulate sensory representations such that they only encode information currently relevant to the organism at minimal cost. This hypothesis has, however, not yet been expressed in a normative computational framework. Here, by building on normative principles of probabilistic inference and efficient coding, we developed a model of dynamic population coding in the visual cortex. By continuously adapting the sensory code to changing demands of the perceptual observer, an attention-like modulation emerges. This modulation can dramatically reduce the amount of neural activity without deteriorating the accuracy of task-specific inferences. Our results suggest that a range of seemingly disparate cortical phenomena such as intrinsic gain modulation, attention-related tuning modulation, and response variability could be manifestations of the same underlying principles, which combine efficient sensory coding with optimal probabilistic inference in dynamic environments.

Motor cortex activity across movement speeds is predicted by network-level strategies for generating muscle activity

Learned movements can be skillfully performed at different paces. What neural strategies produce this flexibility? Can they be predicted and understood by network modeling? We trained monkeys to perform a cycling task at different speeds, and trained artificial recurrent networks to generate the empirical muscle-activity patterns. Network solutions reflected the principle that smooth well-behaved dynamics require low trajectory tangling. Network solutions had a consistent form, which yielded quantitative and qualitative predictions. To evaluate predictions, we analyzed motor cortex activity recorded during the same task. Responses supported the hypothesis that the dominant neural signals reflect not muscle activity, but network-level strategies for generating muscle activity. Single-neuron responses were better accounted for by network activity than by muscle activity. Similarly, neural population trajectories shared their organization not with muscle trajectories, but with network solutions. Thus, cortical activity could be understood based on the need to generate muscle activity via dynamics that allow smooth, robust control over movement speed.

Emerged human-like facial expression representation in a deep convolutional neural network

Recent studies found that the deep convolutional neural networks (DCNNs) trained to recognize facial identities spontaneously learned features that support facial expression recognition, and vice versa. Here, we showed that the self-emerged expression-selective units in a VGG-Face trained for facial identification were tuned to distinct basic expressions and, importantly, exhibited hallmarks of human expression recognition (i.e., facial expression confusion and categorical perception). We then investigated whether the emergence of expression-selective units is attributed to either face-specific experience or domain-general processing by conducting the same analysis on a VGG-16 trained for object classification and an untrained VGG-Face without any visual experience, both having the identical architecture with the pretrained VGG-Face. Although similar expression-selective units were found in both DCNNs, they did not exhibit reliable human-like characteristics of facial expression perception. Together, these findings revealed the necessity of domain-specific visual experience of face identity for the development of facial expression perception, highlighting the contribution of nurture to form human-like facial expression perception.

Selective particle attention: Rapidly and flexibly selecting features for deep reinforcement learning

Deep Reinforcement Learning (RL) is often criticised for being data inefficient and inflexible to changes in task structure. Part of the reason for these issues is that Deep RL typically learns end-to-end using backpropagation, which results in task-specific representations. One approach for circumventing these problems is to apply Deep RL to existing representations that have been learned in a more task-agnostic fashion. However, this only partially solves the problem as the Deep RL algorithm learns a function of all pre-existing representations and is therefore still susceptible to data inefficiency and a lack of flexibility. Biological agents appear to solve this problem by forming internal representations over many tasks and only selecting a subset of these features for decision-making based on the task at hand; a process commonly referred to as selective attention. We take inspiration from selective attention in biological agents and propose a novel algorithm called Selective Particle Attention (SPA), which selects subsets of existing representations for Deep RL. Crucially, these subsets are not learned through backpropagation, which is slow and prone to overfitting, but instead via a particle filter that rapidly and flexibly identifies key subsets of features using only reward feedback. We evaluate SPA on two tasks that involve raw pixel input and dynamic changes to the task structure, and show that it greatly increases the efficiency and flexibility of downstream Deep RL algorithms.

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.

Prediction of Time Series Gene Expression and Structural Analysis of Gene Regulatory Networks Using Recurrent Neural Networks

Methods for time series prediction and classification of gene regulatory networks (GRNs) from gene expression data have been treated separately so far. The recent emergence of attention-based recurrent neural network (RNN) models boosted the interpretability of RNN parameters, making them appealing for the understanding of gene interactions. In this work, we generated synthetic time series gene expression data from a range of archetypal GRNs and we relied on a dual attention RNN to predict the gene temporal dynamics. We show that the prediction is extremely accurate for GRNs with different architectures. Next, we focused on the attention mechanism of the RNN and, using tools from graph theory, we found that its graph properties allow one to hierarchically distinguish different architectures of the GRN. We show that the GRN responded differently to the addition of noise in the prediction by the RNN and we related the noise response to the analysis of the attention mechanism. In conclusion, this work provides a way to understand and exploit the attention mechanism of RNNs and it paves the way to RNN-based methods for time series prediction and inference of GRNs from gene expression data.

Using top-down modulation to optimally balance shared versus separated task representations

Human adaptive behavior requires continually learning and performing a wide variety of tasks, often with very little practice. To accomplish this, it is crucial to separate neural representations of different tasks in order to avoid interference. At the same time, sharing neural representations supports generalization and allows faster learning. Therefore, a crucial challenge is to find an optimal balance between shared versus separated representations. Typically, models of human cognition employ top-down modulatory signals to separate task representations, but there exist surprisingly little systematic computational investigations of how such modulation is best implemented. We identify and systematically evaluate two crucial features of modulatory signals. First, top-down input can be processed in an additive or multiplicative manner. Second, the modulatory signals can be adaptive (learned) or non-adaptive (random). We cross these two features, resulting in four modulation networks which are tested on a variety of input datasets and tasks with different degrees of stimulus-action mapping overlap. The multiplicative adaptive modulation network outperforms all other networks in terms of accuracy. Moreover, this network develops hidden units that optimally share representations between tasks. Specifically, different than the binary approach of currently popular latent state models, it exploits partial overlap between tasks.

Generalized attention-weighted reinforcement learning

In neuroscience, attention has been shown to bidirectionally interact with reinforcement learning (RL) to reduce the dimensionality of task representations, restricting computations to relevant features. In machine learning, despite their popularity, attention mechanisms have seldom been administered to decision-making problems. Here, we leverage a theoretical model from computational neuroscience - the attention-weighted RL (AWRL), defining how humans identify task-relevant features (i.e., that allow value predictions) - to design an applied deep RL paradigm. We formally demonstrate that the conjunction of the self-attention mechanism, widely employed in machine learning, with value function approximation is a general formulation of the AWRL model. To evaluate our agent, we train it on three Atari tasks at different complexity levels, incorporating both task-relevant and irrelevant features. Because the model uses semantic observations, we can uncover not only which features the agent elects to base decisions on, but also how it chooses to compile more complex, relational features from simpler ones. We first show that performance depends in large part on the ability to compile new compound features, rather than mere focus on individual features. In line with neuroscience predictions, self-attention leads to high resiliency to noise (irrelevant features) compared to other benchmark models. Finally, we highlight the importance and separate contributions of both bottom-up and top-down attention in the learning process. Together, these results demonstrate the broader validity of the AWRL framework in complex task scenarios, and illustrate the benefits of a deeper integration between neuroscience-derived models and RL for decision making in machine learning.

Which One? Choosing Favorite Robot After Different Styles of Storytelling and Robots' Conversation

The influence of human-care service robots in human–robot interaction is becoming of great importance, because of the roles that the robots are taking in today’s and future society. Thus, we need to identify how humans can interact, collaborate, and learn from social robots more efficiently. Additionally, it is important to determine the robots’ modalities that can increase the humans’ perceived likeness and knowledge acquisition and enhance human–robot collaboration. The present study aims to identify the optimal social service robots’ modalities that enhance the human learning process and level of enjoyment from the interaction and even attract the humans’ attention to choosing a robot to collaborate with it. Our target group was college students, pre-service teachers. For this purpose, we designed two experiments, each one split in two parts. Both the experiments were between groups, and human participants had the chance to watch the Nao robot performing a storytelling exercise about the history of robots in a museum-educational activity via video annotations. The robot’s modalities were manipulated on its body movements (expressive arm and head gestures) while performing the storytelling, friendly attitude expressions and storytelling, and personality traits. After the robot’s storytelling, participants filled out a knowledge acquisition questionnaire and a self-reported enjoyment level questionnaire. In the second part, we introduce the idea of participants witnessing a conversation between the robots with the different modalities, and they were asked to choose the robot with which they want to collaborate in a similar activity. Results indicated that participants prefer to collaborate with robots with a cheerful personality and expressive body movements. Especially when they were asked to choose between two robots that were cheerful and had expressive body movements, they preferred the one which originally told them the story. Moreover, participants did not prefer to collaborate with a robot with an extremely friendly attitude and storytelling style.

Convolutional Neural Networks as a Model of the Visual System: Past, Present, and Future

Convolutional neural networks (CNNs) were inspired by early findings in the study of biological vision. They have since become successful tools in computer vision and state-of-the-art models of both neural activity and behavior on visual tasks. This review highlights what, in the context of CNNs, it means to be a good model in computational neuroscience and the various ways models can provide insight. Specifically, it covers the origins of CNNs and the methods by which we validate them as models of biological vision. It then goes on to elaborate on what we can learn about biological vision by understanding and experimenting on CNNs and discusses emerging opportunities for the use of CNNs in vision research beyond basic object recognition.

Biological constraints on neural network models of cognitive function

Neural network models are potential tools for improving our understanding of complex brain functions. To address this goal, these models need to be neurobiologically realistic. However, although neural networks have advanced dramatically in recent years and even achieve human-like performance on complex perceptual and cognitive tasks, their similarity to aspects of brain anatomy and physiology is imperfect. Here, we discuss different types of neural models, including localist, auto-associative, hetero-associative, deep and whole-brain networks, and identify aspects under which their biological plausibility can be improved. These aspects range from the choice of model neurons and of mechanisms of synaptic plasticity and learning to implementation of inhibition and control, along with neuroanatomical properties including areal structure and local and long-range connectivity. We highlight recent advances in developing biologically grounded cognitive theories and in mechanistically explaining, on the basis of these brain-constrained neural models, hitherto unaddressed issues regarding the nature, localization and ontogenetic and phylogenetic development of higher brain functions. In closing, we point to possible future clinical applications of brain-constrained modelling.

The Costs and Benefits of Goal-Directed Attention in Deep Convolutional Neural Networks

People deploy top-down, goal-directed attention to accomplish tasks, such as finding lost keys. By tuning the visual system to relevant information sources, object recognition can become more efficient (a benefit) and more biased toward the target (a potential cost). Motivated by selective attention in categorisation models, we developed a goal-directed attention mechanism that can process naturalistic (photographic) stimuli. Our attention mechanism can be incorporated into any existing deep convolutional neural networks (DCNNs). The processing stages in DCNNs have been related to ventral visual stream. In that light, our attentional mechanism incorporates top-down influences from prefrontal cortex (PFC) to support goal-directed behaviour. Akin to how attention weights in categorisation models warp representational spaces, we introduce a layer of attention weights to the mid-level of a DCNN that amplify or attenuate activity to further a goal. We evaluated the attentional mechanism using photographic stimuli, varying the attentional target. We found that increasing goal-directed attention has benefits (increasing hit rates) and costs (increasing false alarm rates). At a moderate level, attention improves sensitivity (i.e. increases d ' ) at only a moderate increase in bias for tasks involving standard images, blended images and natural adversarial images chosen to fool DCNNs. These results suggest that goal-directed attention can reconfigure general-purpose DCNNs to better suit the current task goal, much like PFC modulates activity along the ventral stream. In addition to being more parsimonious and brain consistent, the mid-level attention approach performed better than a standard machine learning approach for transfer learning, namely retraining the final network layer to accommodate the new task.

Image memorability is predicted by discriminability and similarity in different stages of a convolutional neural network

The features of an image can be represented at multiple levels-from its low-level visual properties to high-level meaning. What drives some images to be memorable while others are forgettable? We address this question across two behavioral experiments. In the first, different layers of a convolutional neural network (CNN), which represent progressively higher levels of features, were used to select the images that would be shown to 100 participants through a form of prospective assignment. Here, the discriminability/similarity of an image with others, according to different CNN layers dictated the images presented to different groups, who made a simple indoor versus outdoor judgment for each scene. We found that participants remember more scene images that were selected based on their low-level discriminability or high-level similarity. A second experiment replicated these results in an independent sample of 50 participants, with a different order of postencoding tasks. Together, these experiments provide evidence that both discriminability and similarity, at different visual levels, predict image memorability.

False memories for scenes using the DRM paradigm

People are remarkably good at remembering photographs. To further investigate the nature of the stored representations and the fidelity of human memories, it would be useful to evaluate the visual similarity of stimuli presented in experiments. Here, we explored the possible use of convolutional neural networks (CNN) as a measure of perceptual or representational similarity of visual scenes with respect to visual memory research. In Experiment 1, we presented participants with sets of nine images from the same scene category and tested whether they were able to detect the most distant scene in the image space defined by CNN. Experiment 2 was a visual variant of the Deese-Roediger-McDermott paradigm. We asked participants to remember a set of photographs from the same scene category. The photographs were preselected based on their distance to a particular visual prototype (defined as centroid of the image space). In the recognition test, we observed higher false alarm rates for scenes closer to this visual prototype. Our findings show that the similarity measured by CNN is reflected in human behavior: people can detect odd-one-out scenes or be lured to false alarms with similar stimuli. This method can be used for further studies regarding visual memory for complex scenes.

Attention in Psychology, Neuroscience, and Machine Learning

Attention is the important ability to flexibly control limited computational resources. It has been studied in conjunction with many other topics in neuroscience and psychology including awareness, vigilance, saliency, executive control, and learning. It has also recently been applied in several domains in machine learning. The relationship between the study of biological attention and its use as a tool to enhance artificial neural networks is not always clear. This review starts by providing an overview of how attention is conceptualized in the neuroscience and psychology literature. It then covers several use cases of attention in machine learning, indicating their biological counterparts where they exist. Finally, the ways in which artificial attention can be further inspired by biology for the production of complex and integrative systems is explored.

Beyond the feedforward sweep: feedback computations in the visual cortex

Visual perception involves the rapid formation of a coarse image representation at the onset of visual processing, which is iteratively refined by late computational processes. These early versus late time windows approximately map onto feedforward and feedback processes, respectively. State-of-the-art convolutional neural networks, the main engine behind recent machine vision successes, are feedforward architectures. Their successes and limitations provide critical information regarding which visual tasks can be solved by purely feedforward processes and which require feedback mechanisms. We provide an overview of recent work in cognitive neuroscience and machine vision that highlights the possible role of feedback processes for both visual recognition and beyond. We conclude by discussing important open questions for future research.

A flexible readout mechanism of human sensory representations

Attention can both enhance and suppress cortical sensory representations. However, changing sensory representations can also be detrimental to behavior. Behavioral consequences can be avoided by flexibly changing sensory readout, while leaving the representations unchanged. Here, we asked human observers to attend to and report about either one of two features which control the visibility of motion while making concurrent measurements of cortical activity with BOLD imaging (fMRI). We extend a well-established linking model to account for the relationship between these measurements and find that changes in sensory representation during directed attention are insufficient to explain perceptual reports. Adding a flexible downstream readout is necessary to best explain our data. Such a model implies that observers should be able to recover information about ignored features, a prediction which we confirm behaviorally. Thus, flexible readout is a critical component of the cortical implementation of human adaptive behavior.

How biological attention mechanisms improve task performance in a large-scale visual system model

How does attentional modulation of neural activity enhance performance? Here we use a deep convolutional neural network as a large-scale model of the visual system to address this question. We model the feature similarity gain model of attention, in which attentional modulation is applied according to neural stimulus tuning. Using a variety of visual tasks, we show that neural modulations of the kind and magnitude observed experimentally lead to performance changes of the kind and magnitude observed experimentally. We find that, at earlier layers, attention applied according to tuning does not successfully propagate through the network, and has a weaker impact on performance than attention applied according to values computed for optimally modulating higher areas. This raises the question of whether biological attention might be applied at least in part to optimize function rather than strictly according to tuning. We suggest a simple experiment to distinguish these alternatives.