Slido link for questions for the whole event: https://app.sli.do/event/bayr24RBpGdcveCqzPfdR6/live/questions (or go to slido website and enter 1242 660)

What we pay attention to depends on the context and the task at hand. On the one hand, the prefrontal cortex can modulate how to direct attention outward to the external world. On the other hand, attention to internal states enables metacognition and configuration of internal states using repertoires of memories and skills. I will first discuss ongoing work in which, inspired by the role of attention in affordances and task-sets, we analyze large scale game play data in the XboX 3D game Bleeding Edge in an interpretable way. I will briefly mention ongoing directions including decoding of plans during chess based on eye-tracking. I will conclude with how future models of multi-scale predictive representations could include prefrontal cortical modulation during planning and task performance.

Many deep neural network architectures loosely based on brain networks have recently been shown to replicate neural firing patterns observed in the brain. One of the most exciting and promising novel architectures, the Transformer neural network, was developed without the brain in mind. In this work, we show that transformers, when equipped with recurrent position encodings, replicate the precisely tuned spatial representations of the hippocampal formation; most notably place and grid cells. Furthermore, we show that this result is no surprise since it is closely related to current hippocampal models from neuroscience. We additionally show the transformer version offers dramatic performance gains over the neuroscience version. This work continues to bind computations of artificial and brain networks, offers a novel understanding of the hippocampal-cortical interaction, and suggests how wider cortical areas may perform complex tasks beyond current neuroscience models such as language comprehension.

In robotics, human-robot collaboration works best when robots are responsive to their human partners’ mental states. Human eye gaze has been used as a proxy for one such mental state: attention. While eye gaze can be a useful signal, for example enabling intent prediction, it is also a noisy one. Gaze serves several functions beyond attention, and thus recognizing what people are attending to from their eye gaze is a complex task. In this talk, I will discuss our research on modeling eye gaze to understand human attention in collaborative tasks such as shared manipulation and assisted driving.

When people make sense of the world, they don't only pay attention to what's actually happening. Their mind also takes them to counterfactual worlds of what could have happened. In this talk, I will illustrate how we can use eye-tracking to uncover the human mind's forays into the imaginary. I will show that when people make causal judgments about physical interactions, they don't just look at what actually happens. They mentally simulate what would have happened in relevant counterfactual situations to assess whether the cause made a difference. And when people try to figure out what happened in the past, they mentally simulate the different scenarios that could have led to the outcome. Together these studies illustrate how attention is not only driven by what's out there in the world, but also by what's hidden inside the mind.

We consider the foundations of attention mechanisms in deep neural network architectures and present three main results. First, we provide a systematic taxonomy of all possible attention mechanisms within, or as extensions of, the McCulloch and Pitt standard model into 18 classes depending on the origin type of the attention signal, the target type of the attention signal, and whether the interaction type is additive or multiplicative. Second, using this taxonomy, we identify three key attention mechanisms: output gating, synaptic gating, and multiplexing. Output gating and synaptic gating are extensions of the standard model and all current attention-based architectures, including transformers, use either output gating or synaptic gating, or a combination of both. Third, we develop a theory of attention capacity and derive mathematical results about the capacity of basic attention networks. For example, the output gating of a linear threshold gate of $n$ variables by another linear threshold gate of the same $n$ variables has capacity $2n^2 (1+o(1))$. Perhaps surprisingly, multiplexing attention is used in the proofs of these results. Synaptic and output gating provide computationally efficient extensions of the standard model allowing for {\it sparse} quadratic activation functions. They can also be viewed as primitives enabling the concise collapsing of multiple layers of processing in the standard model.

The debate around the interpretability of attention mechanisms is centered on whether attention scores can be used as a proxy for the relative amounts of signal carried by sub-components of data. We propose to study the interpretability of attention in the context of set machine learning, where each data point is composed of an unordered collection of instances with a global label. For classical multiple-instance-learning problems and simple extensions, there is a well-defined “importance” ground truth that can be leveraged to cast interpretation as a binary classification problem, which we can quantitatively evaluate. By building synthetic datasets over several data modalities, we perform a systematic assessment of attention-based interpretations. We find that attention distributions are indeed often reflective of the relative importance of individual instances, but that silent failures happen where a model will have high classification performance but attention patterns that do not align with expectations. Based on these observations, we propose to use ensembling to minimize the risk of misleading attention-based explanations.

Submit your questions here: https://app.sli.do/event/bayr24RBpGdcveCqzPfdR6/live/questions

Panelists: Megan deBettencourt, Tobias Gerstenberg, Erin Grant, Ida Monennejad, Ramakrishna Vedantam, James Whittington, Cyril Zhang

For the virtual component, chat with poster presenters on our Discord: https://discord.gg/2WP8nM8P

Unconstrained eye gaze estimation using ordinary webcams in smart phones and tablets is immensely useful for many applications. However, current eye gaze estimators are limited in their ability to generalize to a wide range of unconstrained conditions, including, head poses, eye gaze angles and lighting conditions, etc. This is mainly due to the lack of availability of gaze training data in in-the-wild conditions. Notably, eye gaze is a natural form of human communication while humans interact with each other. Visual data (videos or images) containing human interaction are also abundantly available on the internet and are constantly growing as people upload more. Could we leverage visual data containing human interaction to learn unconstrained gaze estimators? In this talk we will describe our foray into addressing this challenging problem. Our findings point to the great potential of human interaction data as a low cost and ubiquitously available source of training data for unconstrained gaze estimators. By lessening the burden of specialized data collection and annotation, we hope to foster greater real-word adoption and proliferation of gaze estimation technology in end-user devices.

Humans and many other animals have an enormous capacity to learn about sensory stimuli and to master new skills. Many of the mechanisms that enable us to learn remain to be understood. One of the greatest challenges of systems neuroscience is to explain how synaptic connections change to support maximally adaptive behaviour. We will provide an overview of factors that determine the change in the strength of synapses. Specifically, we will discuss the influence of attention, neuromodulators and feedback connections in synaptic plasticity and suggest a specific framework, called BrainProp, in which these factors interact to improve the functioning of the entire network.

Much recent work focuses on learning in the brain using presumed biologically plausible variants of supervised learning algorithms. However, the biological plausibility of these approaches is limited, because there is no teacher in the motor cortex that instructs the motor neurons. Instead, learning in the brain usually depends on reward and punishment. BrainProp is a biologically plausible reinforcement learning scheme for deep networks with an any number of layers. The network chooses an action by selecting a unit in the output layer and uses feedback connections to assign credit to the units in lower layers that are responsible for this action. After the choice, the network receives reinforcement so that there is no need for a teacher. We showed how BrainProp is mathematically equivalent to error backpropagation, for one output unit at a time (Pozzi et al., 2020). We illustrate learning of classical and hard image-classification benchmarks (MNIST, CIFAR10, CIFAR100 and Tiny ImageNet) by deep networks. BrainProp achieves an accuracy that is equivalent to that of standard error-backpropagation, and better than other state-of-the-art biologically inspired learning schemes. Additionally, the trial-and-error nature of learning is associated with limited additional training time so that BrainProp is a factor of 1-3.5 times slower. These results provide new insights into how deep learning may be implemented in the brain.

Attention in psychology and neuroscience conceptualizes how the human mind prioritizes information as a result of limited resources. Machine learning systems do not necessarily share the same limits, but implementations of attention have nevertheless proven useful in machine learning across a broad set of domains. Why is this so? I will focus on one aspect: interpretability, which is an ongoing challenge for machine learning systems. I will discuss two different implementations of attention in machine learning that tie closely to conceptualizations of attention in two domains of psychological research. Using these case studies as a starting point, I will discuss the broader strengths and drawbacks of using attention to constrain and interpret how machine learning systems process information. I will end with a problem statement highlighting the need to move away from localized notions to a global view of how attention-like mechanisms modulate information processing in artificial systems.

Attention and eye movements are thought to be a window to the human mind, and have been extensively studied across Neuroscience, Psychology and HCI. However, progress in this area has been severely limited as the underlying methodology relies on specialized hardware that is expensive (upto $30,000) and hard to scale. In this talk, I will present our recent work from Google, which shows that ML applied to smartphone selfie cameras can enable accurate gaze estimation, comparable to state-of-the-art hardware based devices, at 1/100th the cost and without any additional hardware. Via extensive experiments, we show that our smartphone gaze tech can successfully replicate key findings from prior hardware-based eye movement research in Neuroscience and Psychology, across a variety of tasks including traditional oculomotor tasks, saliency analyses on natural images and reading comprehension. We also show that smartphone gaze could enable applications in improved health/wellness, for example, as a potential digital biomarker for detecting mental fatigue. These results show that smartphone-based attention has the potential to unlock advances by scaling eye movement research, and enabling new applications for improved health, wellness and accessibility, such as gaze-based interaction for patients with ALS/stroke that cannot otherwise interact with devices.

The Transformer is an extremely powerful and prominent deep learning architecture. In this work, we challenge the commonly held belief in deep learning that going deeper is better, and show an alternative design approach that is building wider attention Transformers. We demonstrate that wide single layer Transformer models can compete with or outperform deeper ones in a variety of Natural Language Processing (NLP) tasks when both are trained from scratch. The impact of changing the model aspect ratio on Transformers is then studied systematically. This ratio balances the number of layers and the number of attention heads per layer while keeping the total number of attention heads and all other hyperparameters constant. On average, across 4 NLP tasks and 10 attention types, single layer wide models perform 0.3% better than their deep counterparts. We show an in-depth evaluation and demonstrate how wide models require a far smaller memory footprint and can run faster on commodity hardware, in addition, these wider models are also more interpretable. For example, a single layer Transformer on the IMDb byte level text classification has 3.1x faster inference latency on a CPU than its equally accurate deeper counterpart, and is half the size. Our results suggest that the critical direction for building better Transformers for NLP is their width, and that their depth is less relevant.

Attentional gating is a core mechanism supporting behavioral flexibility, but its biological implementation remains uncertain. Gain modulation of neural responses is likely to play a key role, but simply boosting relevant neural responses can be insufficient for improving behavioral outputs, especially in hierarchical circuits. Here we propose a variation of attentional gating that relies on stochastic gain modulation as a dedicated indicator of task relevance. We show that targeted stochastic modulation can be effectively learned and used to fine-tune hierarchical architectures, without reorganization of the underlying circuits. Simulations of such networks demonstrate improvements in learning efficiency and performance in novel tasks, relative to traditional attentional mechanisms based on deterministic gain increases. The effectiveness of this approach relies on the availability of representational bottlenecks in which the task relevant information is localized in small subpopulations of neurons. Overall, this work provides a new mechanism for constructing intelligent systems that can flexibly and robustly adapt to changes in task structure.

Object rearrangement is a challenge for embodied agents because solving these tasks requires generalizing across a combinatorially large set of underlying entities that take the value of object states. Worse, these entities are often unknown and must be inferred from sensory percepts. We present a hierarchical abstraction approach to uncover these underlying entities and achieve combinatorial generalization from unstructured inputs. By constructing a factorized transition graph over clusters of object representations inferred from pixels, we show how to learn a correspondence between intervening on states of entities in the agent's model and acting on objects in the environment. We use this correspondence to develop a method for control that generalizes to different numbers and configurations of objects, which outperforms current offline deep RL methods when evaluated on a set of simulated rearrangement and stacking tasks.

For the virtual component, chat with poster presenters on our Discord: https://discord.gg/2WP8nM8P

Submit your questions here: https://app.sli.do/event/bayr24RBpGdcveCqzPfdR6/live/questions

Panelists: Henny Admoni, David Ha, Brian Kingsbury, John Langford, Shalini De Mello, Vidhya Navalpakkam, Ashish Vaswani

Explainable artificial intelligence is the attempt to elucidate the workings of systems too complex to be directly accessible to human cognition through suitable sideinformation referred to as “explanations”. We present a trainable explanation module for convolutional image classifiers we call bounded logit attention (BLA). The BLA module learns to select a subset of the convolutional feature map for each input instance, which then serves as an explanation for the classifier’s prediction. BLA overcomes several limitations of the instancewise feature selection method “learning to explain” (L2X) introduced by Chen et al. (2018): 1) BLA scales to real-world sized image classification problems, and 2) BLA offers a canonical way to learn explanations of variable size. Due to its modularity BLA lends itself to transfer learning setups and can also be employed as a post-hoc add-on to trained classifiers. Beyond explainability, BLA may serve as a general purpose method for differentiable approximation of subset selection. In a user study we find that BLA explanations are preferred over explanations generated by the popular (Grad-)CAM method (Zhou et al., 2016; Selvaraju et al., 2017).

Language understanding involves processing text with both the grammatical and common-sense contexts of the text fragments. The text "I went to the grocery store and brought home a car" requires both the grammatical context (syntactic) and common-sense context (semantic) to capture the oddity in the sentence. Contextualized text representations learned by Language Models (LMs) are expected to capture a variety of syntactic and semantic contexts from large amounts of training data corpora. Recent work such as ERNIE has shown that infusing the knowledge contexts, where they are available in LMs, results in significant performance gains on General Language Understanding (GLUE) benchmark tasks. However, to our knowledge, no knowledge-aware model has attempted to infuse knowledge through top-down semantics-driven syntactic processing (Eg: Common-sense to Grammatical) and directly operated on the attention mechanism that LMs leverage to learn the data context. We propose a learning framework Top-Down Language Representation (TDLR) to infuse common-sense semantics into LMs. In our implementation, we build on BERT for its rich syntactic knowledge and use the knowledge graphs ConceptNet and WordNet to infuse semantic knowledge.

Modularity and compositionality are promising inductive biases for addressing longstanding problems in machine learning such as better systematic generalization, as well as better transfer and lower forgetting in the context of continual learning. Here we study how attention-based module selection can help achieve compositonal modularity – i.e. decomposition of tasks into meaningful sub-tasks which are tackled by independent architectural entities that we call modules. These sub-tasks must be reusable and the system should be able to learn them without additional supervision. We design a simple experimental setup in which the model is trained to solve mathematical equations with multiple math operations applied sequentially. We study different attention-based module selection strategies, inspired by the principles introduced in the recent literature. We evaluate the method’s ability to learn modules that can recover the underling sub-tasks (operation) used for data generation, as well as the ability to generalize compositionally. We find that meaningful module selection (i.e. routing) is the key to compositional generalization. Further, without access to the privileged information about which part of the input should be used for module selection, the routing component performs poorly for samples that are compositionally out of training distribution. We find that the the main reason for this lies in the routing component, since many of the tested methods perform well OOD if we report the performance of the best performing path at test time. Additionally, we study the role of the number of primitives, the number of training points and bottlenecks for modular specialization.

Transformer networks have seen great success in natural language processing and machine vision, where task objectives such as next word prediction and image classification benefit from nuanced context sensitivity across high-dimensional inputs. However, there is an ongoing debate about how and when transformers can acquire highly structured behavior and achieve systematic generalization. Here, we explore how well a causal transformer can perform a set of algorithmic tasks, including copying, sorting, and hierarchical compositions of these operations. We demonstrate strong generalization to sequences longer than those used in training by replacing the standard positional encoding typically used in transformers with labels arbitrarily paired with items in the sequence. We searched for the layer and head configuration sufficient to solve the task, and performed attention ablation and analyzed encoded representations. We show that two-layer transformers learn generalizable solutions to multi-level problems, develop signs of systematic task decomposition, and exploit shared computation across related tasks. These results provide key insights into the possible structures of within-task and cross-task computations that stacks of attention layers can afford.

Decision-making AI agents are often faced with two important challenges: the depth of the planning horizon, and the branching factor due to having many choices. Hierarchical reinforcement learning methods aim to solve the first problem, by providing shortcuts that skip over multiple time steps. To cope with the breadth, it is desirable to restrict the agent's attention at each step to a reasonable number of possible choices. The concept of affordances (Gibson, 1977) suggests that only certain actions are feasible in certain states. In this work, we first characterize "affordances" as a "hard" attention mechanism that strictly limits the available choices of temporally extended options. We then investigate the role of hard versus soft attention in training data collection, abstract value learning in long-horizon tasks, and handling a growing number of choices. To this end, we present an online, model-free algorithm to learn affordances that can be used to further learn subgoal options. Finally, we identify and empirically demonstrate the settings in which the "paradox of choice" arises, i.e. when having fewer but more meaningful choices improves the learning speed and performance of a reinforcement learning agent.

Polyp segmentation is essential for accelerating the diagnosis of colon cancer. However, it is challenging because of the diverse color, texture, and varying lighting effects of the polyps as well as the subtle difference between the polyp and its surrounding area. To further increase the performance of polyp segmentation, we propose to focus more on the problematic pixels that are harder to predict. To this end, we propose a novel attention module named Fuzzy Attention to focus more on the difficult pixels. Our attention module generates a high attention score for fuzzy pixels usually located near the boundary region. This module can be embedded in any convolution neural network-based backbone network. We embed our module with various backbone networks: Res2Net, ConvNext and Pyramid Vision Transformer and evaluate the models on five polyp segmentation datasets: Kvasir, CVC-300, CVC-ColonDB, CVC-ClinicDB, and ETIS. Our attention module with Res2Net as the backbone network outperforms the reverse attention-based PraNet by a significant amount on all datasets. In addition, our module with PVT as the backbone network achieves state-of-the-art accuracy of 0.937, 0.811, and 0.791 on the CVC-ClinicDB, CVC-ColonDB, and ETIS, respectively, outperforming the latest SA-Net, TransFuse and Polyp-PVT.

The debate around the interpretability of attention mechanisms is centered on whether attention scores can be used as a proxy for the relative amounts of signal carried by sub-components of data. We propose to study the interpretability of attention in the context of set machine learning, where each data point is composed of an unordered collection of instances with a global label. For classical multiple-instance-learning problems and simple extensions, there is a well-defined “importance” ground truth that can be leveraged to cast interpretation as a binary classification problem, which we can quantitatively evaluate. By building synthetic datasets over several data modalities, we perform a systematic assessment of attention-based interpretations. We find that attention distributions are indeed often reflective of the relative importance of individual instances, but that silent failures happen where a model will have high classification performance but attention patterns that do not align with expectations. Based on these observations, we propose to use ensembling to minimize the risk of misleading attention-based explanations.

The Transformer is an extremely powerful and prominent deep learning architecture. In this work, we challenge the commonly held belief in deep learning that going deeper is better, and show an alternative design approach that is building wider attention Transformers. We demonstrate that wide single layer Transformer models can compete with or outperform deeper ones in a variety of Natural Language Processing (NLP) tasks when both are trained from scratch. The impact of changing the model aspect ratio on Transformers is then studied systematically. This ratio balances the number of layers and the number of attention heads per layer while keeping the total number of attention heads and all other hyperparameters constant. On average, across 4 NLP tasks and 10 attention types, single layer wide models perform 0.3% better than their deep counterparts. We show an in-depth evaluation and demonstrate how wide models require a far smaller memory footprint and can run faster on commodity hardware, in addition, these wider models are also more interpretable. For example, a single layer Transformer on the IMDb byte level text classification has 3.1x faster inference latency on a CPU than its equally accurate deeper counterpart, and is half the size. Our results suggest that the critical direction for building better Transformers for NLP is their width, and that their depth is less relevant.

In neuroscience, attention has been associated operationally with enhanced processing of certain sensory inputs depending on external or internal contexts such as cueing, salience, or mental states. In machine learning, attention usually means a multiplicative mechanism whereby the weights in a weighted summation of an input vector are calculated from the input itself or some other context vector. In both scenarios, attention can be conceptualized as a gating mechanism. In this paper, we argue that three-way interactions serve as a normative way to define a gating mechanism in generative probabilistic graphical models. By going a step beyond pairwise interactions, it empowers much more computational efficiency, like a transistor expands possible digital computations. Models with three-way interactions are also easier to scale up and thus to implement biologically. As an example application, we show that a graphical model with three-way interactions provides a normative explanation for divisive normalization in macaque primary visual cortex, an operation adopted widely throughout the cortex to reduce redundancy, save energy, and improve computation.

Object rearrangement is a challenge for embodied agents because solving these tasks requires generalizing across a combinatorially large set of underlying entities that take the value of object states. Worse, these entities are often unknown and must be inferred from sensory percepts. We present a hierarchical abstraction approach to uncover these underlying entities and achieve combinatorial generalization from unstructured inputs. By constructing a factorized transition graph over clusters of object representations inferred from pixels, we show how to learn a correspondence between intervening on states of entities in the agent's model and acting on objects in the environment. We use this correspondence to develop a method for control that generalizes to different numbers and configurations of objects, which outperforms current offline deep RL methods when evaluated on a set of simulated rearrangement and stacking tasks.

Cross-modal attention is widely used in multimodal learning to fuse information from two modalities. However, most existing models only assimilate cross-modal attention indirectly by relying on end-to-end learning and do not directly improve the attention mechanisms. In this paper, we propose a methodology for directly enhancing cross-modal attention by utilizing object-detection models for vision-and-language tasks that deal with image and text information. We used the mask of the detected objects obtained by the detection model as a pseudo label, and we added a loss between the attention map of the multimodal learning model and the pseudo label. The proposed methodology drastically improves the performance of the baseline model across all performance metrics in various popular datasets for the image-captioning task. Moreover, our highly scalable methodology can be applied to any multimodal task in terms of vision-and-language.

Imbuing robots with human-levels of intelligence is a longstanding goal of AI research.A critical aspect of human-level intelligence is spatial reasoning. Spatial reasoning requires a robot to reason about relationships among objects in an environment to estimate the positions of unseen objects. In this work, we introduced a novel graph attention approach for predicting the locations of query objects in partially observable environments. We found that our approach achieved state of the art results on object location prediction tasks. Then, we evaluated our approach on never before seen objects, and we observed zero-shot generalization to estimate the positions of new object types.

With the growing adoption of deep learning for on-device TinyML applications, there has been an ever-increasing demand for more efficient neural network backbones optimized for the edge. Recently, the introduction of attention condenser networks have resulted in low-footprint, highly-efficient, self-attention neural networks that strike a strong balance between accuracy and speed. In this study, we introduce a new faster attention condenser design called double-condensing attention condensers that enable more condensed feature embedding. We further employ a machine-driven design exploration strategy that imposes best practices design constraints for greater efficiency and robustness to produce the macro-micro architecture constructs of the backbone. The resulting backbone (which we name \textbf{AttendNeXt}) achieves significantly higher inference throughput on an embedded ARM processor when compared to several other state-of-the-art efficient backbones ($>10\times$ faster than FB-Net C at higher accuracy and speed and $>10\times$ faster than MobileOne-S1 at smaller size) while having a small model size ($>1.37\times$ smaller than MobileNetv3-L at higher accuracy and speed) and strong accuracy (1.1\% higher top-1 accuracy than MobileViT XS on ImageNet at higher speed). These promising results demonstrate that exploring different efficient architecture designs and self-attention mechanisms can lead to interesting new building blocks for TinyML applications.

Attentional gating is a core mechanism supporting behavioral flexibility, but its biological implementation remains uncertain. Gain modulation of neural responses is likely to play a key role, but simply boosting relevant neural responses can be insufficient for improving behavioral outputs, especially in hierarchical circuits. Here we propose a variation of attentional gating that relies on stochastic gain modulation as a dedicated indicator of task relevance. We show that targeted stochastic modulation can be effectively learned and used to fine-tune hierarchical architectures, without reorganization of the underlying circuits. Simulations of such networks demonstrate improvements in learning efficiency and performance in novel tasks, relative to traditional attentional mechanisms based on deterministic gain increases. The effectiveness of this approach relies on the availability of representational bottlenecks in which the task relevant information is localized in small subpopulations of neurons. Overall, this work provides a new mechanism for constructing intelligent systems that can flexibly and robustly adapt to changes in task structure.

Transformer-based models have gained large popularity and demonstrated promising results in long-term time-series forecasting in recent years. In addition to learning attention in time domain, recent works also explore learning attention in frequency domains (e.g., Fourier domain, wavelet domain), given that seasonal patterns can be better captured in these domains. In this work, we seek to understand the relationships between attention models in different time and frequency domains. Theoretically, we show that attention models in different domains are equivalent under linear conditions (i.e., linear kernel to attention scores). Empirically, we analyze how attention models of different domains show different behaviors through various synthetic experiments with seasonality, trend and noises, with emphasis on the role of softmax operation therein. Both these theoretical and empirical analyses motivate us to propose a new method: TDformer (Trend Decomposition Transformer), that first applies seasonal-trend decomposition, and then additively combines an MLP which predicts the trend component with Fourier attention which predicts the seasonal component to obtain the final prediction. Extensive experiments on benchmark time-series forecasting datasets demonstrate that TDformer achieves state-of-the-art performance against existing attention-based models.

Modern computational systems have an unprecedented ability to detect, leverage and influence human attention. Prior work identified user engagement and dwell time as two key metrics of attention in digital environments, but these metrics have yet to be integrated into a unified model that can advance the theory and practice of digital attention. We draw on work from cognitive science, digital advertising, and AI to propose a two-stage model of attention for social media environments that disentangles engagement and dwell. In an online experiment, we show that attention operates differently in these two stages and find clear evidence of dissociation: when dwelling on posts (Stage 1), users attend more to sensational than credible content, but when deciding whether to engage with content (Stage 2), users attend more to credible than sensational content. These findings have implications for the design and development of computational systems that measure and model human attention, such as newsfeed algorithms on social media.

Attention mechanisms have recently demonstrated impressive performance on a range of NLP tasks, and attention scores are often used as a proxy for model explainability. However, there is a debate on whether attention weights can, in fact, be used to identify the most important inputs to a model. We approach this question from an information theoretic perspective by measuring the mutual information between the model output and the hidden states. From extensive experiments, we draw the following conclusions: (i) Additive and Deep attention mechanisms are likely to be better at preserving the information between the hidden states and the model output (compared to Scaled Dot-product); (ii) ablation studies indicate that Additive attention can actively learn to explain the importance of its input hidden representations; (iii) when attention values are nearly the same, the rank order of attention values is not consistent with the rank order of the mutual information (iv) Using Gumbel-Softmax with a temperature lower than one, tends to produce a more skewed attention score distribution compared to softmax and hence is a better choice for explainable design; (v) some building blocks are better at preserving the correlation between the ordered list of mutual information and attention weights order (for eg. the combination of BiLSTM encoder and Additive attention). Our findings indicate that attention mechanisms do have the potential to function as a shortcut to model explanations when they are carefully combined with other model elements.

We consider the foundations of attention mechanisms in deep neural network architectures and present three main results. First, we provide a systematic taxonomy of all possible attention mechanisms within, or as extensions of, the McCulloch and Pitt standard model into 18 classes depending on the origin type of the attention signal, the target type of the attention signal, and whether the interaction type is additive or multiplicative. Second, using this taxonomy, we identify three key attention mechanisms: output gating, synaptic gating, and multiplexing. Output gating and synaptic gating are extensions of the standard model and all current attention-based architectures, including transformers, use either output gating or synaptic gating, or a combination of both. Third, we develop a theory of attention capacity and derive mathematical results about the capacity of basic attention networks. For example, the output gating of a linear threshold gate of $n$ variables by another linear threshold gate of the same $n$ variables has capacity $2n^2 (1+o(1))$. Perhaps surprisingly, multiplexing attention is used in the proofs of these results. Synaptic and output gating provide computationally efficient extensions of the standard model allowing for {\it sparse} quadratic activation functions. They can also be viewed as primitives enabling the concise collapsing of multiple layers of processing in the standard model.

Slot attention is a successful method for object-centric modeling with images and videos for tasks like unsupervised object discovery. However, set-equivariance limits its ability to perform tiebreaking, which makes distinguishing similar structures difficult – a task crucial for vision problems. To fix this, we cast cross-attention in slot attention as an optimal transport (OT) problem that has solutions with the desired tiebreaking properties. We then propose an entropy minimization module that combines the tiebreaking properties of unregularized OT with the speed of regularized OT. We evaluate our method on CLEVR object detection and observe significant improvements from 53% to 91% on a strict average precision metric.