Most models of visual attention are aimed at predicting either top-down or bottom-up control, as studied using different visual search and free-viewing tasks. We propose Human Attention Transformer (HAT), a single model predicting both forms of attention control. HAT is the new state-of-the-art (SOTA) in predicting the scanpath of fixations made during target-present and target-absent search, and matches or exceeds SOTA in the prediction of taskless free-viewing fixation scanpaths. HAT achieves this new SOTA by using a novel transformer-based architecture and a simplified foveated retina that collectively create a spatio-temporal awareness akin to the dynamic visual working memory of humans. Unlike previous methods that rely on a coarse grid of fixation cells and experience information loss due to fixation discretization, HAT features a dense-prediction architecture and outputs a dense heatmap for each fixation, thus avoiding discretizing fixations. HAT sets a new standard in computational attention, which emphasizes both effectiveness and generality. HAT's demonstrated scope and applicability will likely inspire the development of new attention models that can better predict human behavior in various attention-demanding scenarios.
相關內容
Achieving high-quality semantic segmentation predictions using only image-level labels enables a new level of real-world applicability. Although state-of-the-art networks deliver reliable predictions, the amount of handcrafted pixel-wise annotations to enable these results are not feasible in many real-world applications. Hence, several works have already targeted this bottleneck, using classifier-based networks like Class Activation Maps~\cite{CAM} (CAMs) as a base. Addressing CAM's weaknesses of fuzzy borders and incomplete predictions, state-of-the-art approaches rely only on adding regulations to the classifier loss or using pixel-similarity-based refinement after the fact. We propose a framework that introduces an additional module using object perimeters for improved saliency. We define object perimeter information as the line separating the object and background. Our new PerimeterFit module will be applied to pre-refine the CAM predictions before using the pixel-similarity-based network. In this way, our PerimeterFit increases the quality of the CAM prediction while simultaneously improving the false negative rate. We investigated a wide range of state-of-the-art unsupervised semantic segmentation networks and edge detection techniques to create useful perimeter maps, which enable our framework to predict object locations with sharper perimeters. We achieved up to 1.5% improvement over frameworks without our PerimeterFit module. We conduct an exhaustive analysis to illustrate that SILOP enhances existing state-of-the-art frameworks for image-level-based semantic segmentation. The framework is open-source and accessible online at //github.com/ErikOstrowski/SILOP.
Human motion prediction aims to forecast an upcoming pose sequence given a past human motion trajectory. To address the problem, in this work we propose FreqMRN, a human motion prediction framework that takes into account both the kinematic structure of the human body and the temporal smoothness nature of motion. Specifically, FreqMRN first generates a fixed-size motion history summary using a motion attention module, which helps avoid inaccurate motion predictions due to excessively long motion inputs. Then, supervised by the proposed spatial-temporal-aware, velocity-aware and global-smoothness-aware losses, FreqMRN iteratively refines the predicted motion though the proposed motion refinement module, which converts motion representations back and forth between pose space and frequency space. We evaluate FreqMRN on several standard benchmark datasets, including Human3.6M, AMASS and 3DPW. Experimental results demonstrate that FreqMRN outperforms previous methods by large margins for both short-term and long-term predictions, while demonstrating superior robustness.
Current state-of-the-art methods for automatic synthetic speech evaluation are based on MOS prediction neural models. Such MOS prediction models include MOSNet and LDNet that use spectral features as input, and SSL-MOS that relies on a pretrained self-supervised learning model that directly uses the speech signal as input. In modern high-quality neural TTS systems, prosodic appropriateness with regard to the spoken content is a decisive factor for speech naturalness. For this reason, we propose to include prosodic and linguistic features as additional inputs in MOS prediction systems, and evaluate their impact on the prediction outcome. We consider phoneme level F0 and duration features as prosodic inputs, as well as Tacotron encoder outputs, POS tags and BERT embeddings as higher-level linguistic inputs. All MOS prediction systems are trained on SOMOS, a neural TTS-only dataset with crowdsourced naturalness MOS evaluations. Results show that the proposed additional features are beneficial in the MOS prediction task, by improving the predicted MOS scores' correlation with the ground truths, both at utterance-level and system-level predictions.
In recent years, several reaction templates-based and template-free approaches have been reported for single-step retrosynthesis prediction. Even though many of these approaches perform well from traditional data-driven metrics standpoint, there is a disconnect between model architectures used and underlying chemistry principles governing retrosynthesis. Here, we propose a novel chemistry-aware retrosynthesis prediction framework that combines powerful data-driven models with chemistry knowledge. We report a tree-to-sequence transformer architecture based on hierarchical SMILES grammar trees as input containing underlying chemistry information that is otherwise ignored by models based on purely SMILES-based representations. The proposed framework, grammar-based molecular attention tree transformer (G-MATT), achieves significant performance improvements compared to baseline retrosynthesis models. G-MATT achieves a top-1 accuracy of 51% (top-10 accuracy of 79.1%), invalid rate of 1.5%, and bioactive similarity rate of 74.8%. Further analyses based on attention maps demonstrate G-MATT's ability to preserve chemistry knowledge without having to use extremely complex model architectures.
Humans have the natural ability to recognize actions even if the objects involved in the action or the background are changed. Humans can abstract away the action from the appearance of the objects and their context which is referred to as compositionality of actions. Compositional action recognition deals with imparting human-like compositional generalization abilities to action-recognition models. In this regard, extracting the interactions between humans and objects forms the basis of compositional understanding. These interactions are not affected by the appearance biases of the objects or the context. But the context provides additional cues about the interactions between things and stuff. Hence we need to infuse context into the human-object interactions for compositional action recognition. To this end, we first design a spatial-temporal interaction encoder that captures the human-object (things) interactions. The encoder learns the spatio-temporal interaction tokens disentangled from the background context. The interaction tokens are then infused with contextual information from the video tokens to model the interactions between things and stuff. The final context-infused spatio-temporal interaction tokens are used for compositional action recognition. We show the effectiveness of our interaction-centric approach on the compositional Something-Else dataset where we obtain a new state-of-the-art result of 83.8% top-1 accuracy outperforming recent important object-centric methods by a significant margin. Our approach of explicit human-object-stuff interaction modeling is effective even for standard action recognition datasets such as Something-Something-V2 and Epic-Kitchens-100 where we obtain comparable or better performance than state-of-the-art.
Modern neural network training relies heavily on data augmentation for improved generalization. After the initial success of label-preserving augmentations, there has been a recent surge of interest in label-perturbing approaches, which combine features and labels across training samples to smooth the learned decision surface. In this paper, we propose a new augmentation method that leverages the first and second moments extracted and re-injected by feature normalization. We replace the moments of the learned features of one training image by those of another, and also interpolate the target labels. As our approach is fast, operates entirely in feature space, and mixes different signals than prior methods, one can effectively combine it with existing augmentation methods. We demonstrate its efficacy across benchmark data sets in computer vision, speech, and natural language processing, where it consistently improves the generalization performance of highly competitive baseline networks.
Current state-of-the-art semantic role labeling (SRL) uses a deep neural network with no explicit linguistic features. However, prior work has shown that gold syntax trees can dramatically improve SRL decoding, suggesting the possibility of increased accuracy from explicit modeling of syntax. In this work, we present linguistically-informed self-attention (LISA): a neural network model that combines multi-head self-attention with multi-task learning across dependency parsing, part-of-speech tagging, predicate detection and SRL. Unlike previous models which require significant pre-processing to prepare linguistic features, LISA can incorporate syntax using merely raw tokens as input, encoding the sequence only once to simultaneously perform parsing, predicate detection and role labeling for all predicates. Syntax is incorporated by training one attention head to attend to syntactic parents for each token. Moreover, if a high-quality syntactic parse is already available, it can be beneficially injected at test time without re-training our SRL model. In experiments on CoNLL-2005 SRL, LISA achieves new state-of-the-art performance for a model using predicted predicates and standard word embeddings, attaining 2.5 F1 absolute higher than the previous state-of-the-art on newswire and more than 3.5 F1 on out-of-domain data, nearly 10% reduction in error. On ConLL-2012 English SRL we also show an improvement of more than 2.5 F1. LISA also out-performs the state-of-the-art with contextually-encoded (ELMo) word representations, by nearly 1.0 F1 on news and more than 2.0 F1 on out-of-domain text.
Most existing works in visual question answering (VQA) are dedicated to improving the accuracy of predicted answers, while disregarding the explanations. We argue that the explanation for an answer is of the same or even more importance compared with the answer itself, since it makes the question and answering process more understandable and traceable. To this end, we propose a new task of VQA-E (VQA with Explanation), where the computational models are required to generate an explanation with the predicted answer. We first construct a new dataset, and then frame the VQA-E problem in a multi-task learning architecture. Our VQA-E dataset is automatically derived from the VQA v2 dataset by intelligently exploiting the available captions. We have conducted a user study to validate the quality of explanations synthesized by our method. We quantitatively show that the additional supervision from explanations can not only produce insightful textual sentences to justify the answers, but also improve the performance of answer prediction. Our model outperforms the state-of-the-art methods by a clear margin on the VQA v2 dataset.
Many natural language processing tasks solely rely on sparse dependencies between a few tokens in a sentence. Soft attention mechanisms show promising performance in modeling local/global dependencies by soft probabilities between every two tokens, but they are not effective and efficient when applied to long sentences. By contrast, hard attention mechanisms directly select a subset of tokens but are difficult and inefficient to train due to their combinatorial nature. In this paper, we integrate both soft and hard attention into one context fusion model, "reinforced self-attention (ReSA)", for the mutual benefit of each other. In ReSA, a hard attention trims a sequence for a soft self-attention to process, while the soft attention feeds reward signals back to facilitate the training of the hard one. For this purpose, we develop a novel hard attention called "reinforced sequence sampling (RSS)", selecting tokens in parallel and trained via policy gradient. Using two RSS modules, ReSA efficiently extracts the sparse dependencies between each pair of selected tokens. We finally propose an RNN/CNN-free sentence-encoding model, "reinforced self-attention network (ReSAN)", solely based on ReSA. It achieves state-of-the-art performance on both Stanford Natural Language Inference (SNLI) and Sentences Involving Compositional Knowledge (SICK) datasets.
Semantic Role Labeling (SRL) is believed to be a crucial step towards natural language understanding and has been widely studied. Recent years, end-to-end SRL with recurrent neural networks (RNN) has gained increasing attention. However, it remains a major challenge for RNNs to handle structural information and long range dependencies. In this paper, we present a simple and effective architecture for SRL which aims to address these problems. Our model is based on self-attention which can directly capture the relationships between two tokens regardless of their distance. Our single model achieves F$_1=83.4$ on the CoNLL-2005 shared task dataset and F$_1=82.7$ on the CoNLL-2012 shared task dataset, which outperforms the previous state-of-the-art results by $1.8$ and $1.0$ F$_1$ score respectively. Besides, our model is computationally efficient, and the parsing speed is 50K tokens per second on a single Titan X GPU.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191