对于沉浸式应用，匹配视觉同行的双耳发电是对虚拟环境中的人们带来有意义的体验至关重要。最近的作品已经显示了使用神经网络来使用2D视觉信息作为指导来使用Mono音频来合成双耳音频。通过使用3D视觉信息引导音频并在波形域中操作来扩展该方法可以允许虚拟音频场景的更准确的Auratization。在本文中，我们提供了一个多模态深入学习模型的点，它使用3D点云场景从单声道音频生成双耳版本。具体地，Point2Sound由具有3D稀疏卷积的视觉网络组成，其从点云场景中提取视觉特征来调节操作在波形域中的音频网络，以合成双耳网络。实验结果表明，3D视觉信息可以成功引导双模深度学习模型的双耳合成任务。此外，我们还调查了不同的丢失函数和3D点云属性，显示直接预测完整的双耳信号并使用RGB深度特征增加了我们所提出的模型的性能。

We derive a set of causal deep neural networks whose architectures are a consequence of tensor (multilinear) factor analysis. Forward causal questions are addressed with a neural network architecture composed of causal capsules and a tensor transformer. The former estimate a set of latent variables that represent the causal factors, and the latter governs their interaction. Causal capsules and tensor transformers may be implemented using shallow autoencoders, but for a scalable architecture we employ block algebra and derive a deep neural network composed of a hierarchy of autoencoders. An interleaved kernel hierarchy preprocesses the data resulting in a hierarchy of kernel tensor factor models. Inverse causal questions are addressed with a neural network that implements multilinear projection and estimates the causes of effects. As an alternative to aggressive bottleneck dimension reduction or regularized regression that may camouflage an inherently underdetermined inverse problem, we prescribe modeling different aspects of the mechanism of data formation with piecewise tensor models whose multilinear projections are well-defined and produce multiple candidate solutions. Our forward and inverse neural network architectures are suitable for asynchronous parallel computation.

The existing methods for video anomaly detection mostly utilize videos containing identifiable facial and appearance-based features. The use of videos with identifiable faces raises privacy concerns, especially when used in a hospital or community-based setting. Appearance-based features can also be sensitive to pixel-based noise, straining the anomaly detection methods to model the changes in the background and making it difficult to focus on the actions of humans in the foreground. Structural information in the form of skeletons describing the human motion in the videos is privacy-protecting and can overcome some of the problems posed by appearance-based features. In this paper, we present a survey of privacy-protecting deep learning anomaly detection methods using skeletons extracted from videos. We present a novel taxonomy of algorithms based on the various learning approaches. We conclude that skeleton-based approaches for anomaly detection can be a plausible privacy-protecting alternative for video anomaly detection. Lastly, we identify major open research questions and provide guidelines to address them.

Several self-supervised representation learning methods have been proposed for reinforcement learning (RL) with rich observations. For real-world applications of RL, recovering underlying latent states is crucial, particularly when sensory inputs contain irrelevant and exogenous information. In this work, we study how information bottlenecks can be used to construct latent states efficiently in the presence of task-irrelevant information. We propose architectures that utilize variational and discrete information bottlenecks, coined as RepDIB, to learn structured factorized representations. Exploiting the expressiveness bought by factorized representations, we introduce a simple, yet effective, bottleneck that can be integrated with any existing self-supervised objective for RL. We demonstrate this across several online and offline RL benchmarks, along with a real robot arm task, where we find that compressed representations with RepDIB can lead to strong performance improvements, as the learned bottlenecks help predict only the relevant state while ignoring irrelevant information.

We propose reconstruction probing, a new analysis method for contextualized representations based on reconstruction probabilities in masked language models (MLMs). This method relies on comparing the reconstruction probabilities of tokens in a given sequence when conditioned on the representation of a single token that has been fully contextualized and when conditioned on only the decontextualized lexical prior of the model. This comparison can be understood as quantifying the contribution of contextualization towards reconstruction -- the difference in the reconstruction probabilities can only be attributed to the representational change of the single token induced by contextualization. We apply this analysis to three MLMs and find that contextualization boosts reconstructability of tokens that are close to the token being reconstructed in terms of linear and syntactic distance. Furthermore, we extend our analysis to finer-grained decomposition of contextualized representations, and we find that these boosts are largely attributable to static and positional embeddings at the input layer.

Diffusion models have achieved justifiable popularity by attaining state-of-the-art performance in generating realistic objects from seemingly arbitrarily complex data distributions, including when conditioning generation on labels. Unfortunately, however, their iterative nature renders them very computationally inefficient during the sampling process. For the multi-class conditional generation problem, we propose a novel, structurally unique framework of diffusion models which are hierarchically branched according to the inherent relationships between classes. In this work, we demonstrate that branched diffusion models offer major improvements in efficiently generating samples from multiple classes. We also showcase several other advantages of branched diffusion models, including ease of extension to novel classes in a continual-learning setting, and a unique interpretability that offers insight into these generative models. Branched diffusion models represent an alternative paradigm to their traditional linear counterparts, and can have large impacts in how we use diffusion models for efficient generation, online learning, and scientific discovery.

The widely studied task of Natural Language Inference (NLI) requires a system to recognize whether one piece of text is textually entailed by another, i.e. whether the entirety of its meaning can be inferred from the other. In current NLI datasets and models, textual entailment relations are typically defined on the sentence- or paragraph-level. However, even a simple sentence often contains multiple propositions, i.e. distinct units of meaning conveyed by the sentence. As these propositions can carry different truth values in the context of a given premise, we argue for the need to recognize the textual entailment relation of each proposition in a sentence individually. We propose PropSegmEnt, a corpus of over 35K propositions annotated by expert human raters. Our dataset structure resembles the tasks of (1) segmenting sentences within a document to the set of propositions, and (2) classifying the entailment relation of each proposition with respect to a different yet topically-aligned document, i.e. documents describing the same event or entity. We establish strong baselines for the segmentation and entailment tasks. Through case studies on summary hallucination detection and document-level NLI, we demonstrate that our conceptual framework is potentially useful for understanding and explaining the compositionality of NLI labels.

People living with dementia often exhibit behavioural and psychological symptoms of dementia that can put their and others' safety at risk. Existing video surveillance systems in long-term care facilities can be used to monitor such behaviours of risk to alert the staff to prevent potential injuries or death in some cases. However, these behaviours of risk events are heterogeneous and infrequent in comparison to normal events. Moreover, analyzing raw videos can also raise privacy concerns. In this paper, we present two novel privacy-protecting video-based anomaly detection approaches to detect behaviours of risks in people with dementia. We either extracted body pose information as skeletons and use semantic segmentation masks to replace multiple humans in the scene with their semantic boundaries. Our work differs from most existing approaches for video anomaly detection that focus on appearance-based features, which can put the privacy of a person at risk and is also susceptible to pixel-based noise, including illumination and viewing direction. We used anonymized videos of normal activities to train customized spatio-temporal convolutional autoencoders and identify behaviours of risk as anomalies. We show our results on a real-world study conducted in a dementia care unit with patients with dementia, containing approximately 21 hours of normal activities data for training and 9 hours of data containing normal and behaviours of risk events for testing. We compared our approaches with the original RGB videos and obtained an equivalent area under the receiver operating characteristic curve performance of 0.807 for the skeleton-based approach and 0.823 for the segmentation mask-based approach. This is one of the first studies to incorporate privacy for the detection of behaviours of risks in people with dementia.

Language models have recently achieved strong performance across a wide range of NLP benchmarks. However, unlike benchmarks, real world tasks are often poorly specified, and agents must deduce the user's intended behavior from a combination of context, instructions, and examples. We investigate how both humans and models behave in the face of such task ambiguity by proposing AmbiBench, a new benchmark of six ambiguously-specified classification tasks. We evaluate humans and models on AmbiBench by seeing how well they identify the intended task using 1) instructions with varying degrees of ambiguity, and 2) different numbers of labeled examples. We find that the combination of model scaling (to 175B parameters) and training with human feedback data enables models to approach or exceed the accuracy of human participants across tasks, but that either one alone is not sufficient. In addition, we show how to dramatically improve the accuracy of language models trained without large-scale human feedback training by finetuning on a small number of ambiguous in-context examples, providing a promising direction for teaching models to generalize well in the face of ambiguity.

Despite their impressive performance on diverse tasks, large language models (LMs) still struggle with tasks requiring rich world knowledge, implying the limitations of relying solely on their parameters to encode a wealth of world knowledge. This paper aims to understand LMs' strengths and limitations in memorizing factual knowledge, by conducting large-scale knowledge probing experiments of 10 models and 4 augmentation methods on PopQA, our new open-domain QA dataset with 14k questions. We find that LMs struggle with less popular factual knowledge, and that scaling fails to appreciably improve memorization of factual knowledge in the tail. We then show that retrieval-augmented LMs largely outperform orders of magnitude larger LMs, while unassisted LMs remain competitive in questions about high-popularity entities. Based on those findings, we devise a simple, yet effective, method for powerful and efficient retrieval-augmented LMs, which retrieves non-parametric memories only when necessary. Experimental results show that this significantly improves models' performance while reducing the inference costs.