Transformer has achieved remarkable success in language, image, and speech processing. Recently, various efficient attention architectures have been proposed to improve transformer's efficiency while largely preserving its efficacy, especially in modeling long sequences. A widely-used benchmark to test these efficient methods' capability on long-range modeling is Long Range Arena (LRA). However, LRA only focuses on the standard bidirectional (or noncausal) self attention, and completely ignores cross attentions and unidirectional (or causal) attentions, which are equally important to downstream applications. Although designing cross and causal variants of an attention method is straightforward for vanilla attention, it is often challenging for efficient attentions with subquadratic time and memory complexity. In this paper, we propose Comprehensive Attention Benchmark (CAB) under a fine-grained attention taxonomy with four distinguishable attention patterns, namely, noncausal self, causal self, noncausal cross, and causal cross attentions. CAB collects seven real-world tasks from different research areas to evaluate efficient attentions under the four attention patterns. Among these tasks, CAB validates efficient attentions in eight backbone networks to show their generalization across neural architectures. We conduct exhaustive experiments to benchmark the performances of nine widely-used efficient attention architectures designed with different philosophies on CAB. Extensive experimental results also shed light on the fundamental problems of efficient attentions, such as efficiency length against vanilla attention, performance consistency across attention patterns, the benefit of attention mechanisms, and interpolation/extrapolation on long-context language modeling.
Reinforcement Learning (RL) algorithms are often known for sample inefficiency and difficult generalization. Recently, Unsupervised Environment Design (UED) emerged as a new paradigm for zero-shot generalization by simultaneously learning a task distribution and agent policies on the sampled tasks. This is a non-stationary process where the task distribution evolves along with agent policies, creating an instability over time. While past works demonstrated the potential of such approaches, sampling effectively from the task space remains an open challenge, bottlenecking these approaches. To this end, we introduce CLUTR: a novel curriculum learning algorithm that decouples task representation and curriculum learning into a two-stage optimization. It first trains a recurrent variational autoencoder on randomly generated tasks to learn a latent task manifold. Next, a teacher agent creates a curriculum by maximizing a minimax REGRET-based objective on a set of latent tasks sampled from this manifold. By keeping the task manifold fixed, we show that CLUTR successfully overcomes the non-stationarity problem and improves stability. Our experimental results show CLUTR outperforms PAIRED, a principled and popular UED method, in terms of generalization and sample efficiency in the challenging CarRacing and navigation environments: showing an 18x improvement on the F1 CarRacing benchmark. CLUTR also performs comparably to the non-UED state-of-the-art for CarRacing, outperforming it in nine of the 20 tracks. CLUTR also achieves a 33% higher solved rate than PAIRED on a set of 18 out-of-distribution navigation tasks.
Real-time 3D mapping is a critical component in many important applications today including robotics, AR/VR, and 3D visualization. 3D mapping involves continuously fusing depth maps obtained from depth sensors in phones, robots, and autonomous vehicles into a single 3D representative model of the scene. Many important applications, e.g., global path planning and trajectory generation in micro aerial vehicles, require the construction of large maps at high resolutions. In this work, we identify mapping, i.e., construction and updates of 3D maps to be a critical bottleneck in these applications. The memory required and access times of these maps limit the size of the environment and the resolution with which the environment can be feasibly mapped, especially in resource constrained environments such as autonomous robot platforms and portable devices. To address this challenge, we propose VoxelCache: a hardware-software technique to accelerate map data access times in 3D mapping applications. We observe that mapping applications typically access voxels in the map that are spatially co-located to each other. We leverage this temporal locality in voxel accesses to cache indices to blocks of voxels to enable quick lookup and avoid expensive access times. We evaluate VoxelCache on popularly used mapping and reconstruction applications on both GPUs and CPUs. We demonstrate an average speedup of 1.47X (up to 1.66X) and 1.79X (up to 1.91X) on CPUs and GPUs respectively.
Recently, multi-channel speech enhancement has drawn much interest due to the use of spatial information to distinguish target speech from interfering signal. To make full use of spatial information and neural network based masking estimation, we propose a multi-channel denoising neural network -- Spatial DCCRN. Firstly, we extend S-DCCRN to multi-channel scenario, aiming at performing cascaded sub-channel and full-channel processing strategy, which can model different channels separately. Moreover, instead of only adopting multi-channel spectrum or concatenating first-channel's magnitude and IPD as the model's inputs, we apply an angle feature extraction module (AFE) to extract frame-level angle feature embeddings, which can help the model to apparently perceive spatial information. Finally, since the phenomenon of residual noise will be more serious when the noise and speech exist in the same time frequency (TF) bin, we particularly design a masking and mapping filtering method to substitute the traditional filter-and-sum operation, with the purpose of cascading coarsely denoising, dereverberation and residual noise suppression. The proposed model, Spatial-DCCRN, has surpassed EaBNet, FasNet as well as several competitive models on the L3DAS22 Challenge dataset. Not only the 3D scenario, Spatial-DCCRN outperforms state-of-the-art (SOTA) model MIMO-UNet by a large margin in multiple evaluation metrics on the multi-channel ConferencingSpeech2021 Challenge dataset. Ablation studies also demonstrate the effectiveness of different contributions.
Developing robot controllers in a simulated environment is advantageous but transferring the controllers to the target environment presents challenges, often referred to as the "sim-to-real gap". We present a method for continuous improvement of modeling and control after deploying the robot to a dynamically-changing target environment. We develop a differentiable physics simulation framework that performs online system identification and optimal control simultaneously, using the incoming observations from the target environment in real time. To ensure robust system identification against noisy observations, we devise an algorithm to assess the confidence of our estimated parameters, using numerical analysis of the dynamic equations. To ensure real-time optimal control, we adaptively schedule the optimization window in the future so that the optimized actions can be replenished faster than they are consumed, while staying as up-to-date with new sensor information as possible. The constant re-planning based on a constantly improved model allows the robot to swiftly adapt to the changing environment and utilize real-world data in the most sample-efficient way. Thanks to a fast differentiable physics simulator, the optimization for both system identification and control can be solved efficiently for robots operating in real time. We demonstrate our method on a set of examples in simulation and show that our results are favorable compared to baseline methods.
Recently, continual learning has received a lot of attention. One of the significant problems is the occurrence of \emph{concept drift}, which consists of changing probabilistic characteristics of the incoming data. In the case of the classification task, this phenomenon destabilizes the model's performance and negatively affects the achieved prediction quality. Most current methods apply statistical learning and similarity analysis over the raw data. However, similarity analysis in streaming data remains a complex problem due to time limitation, non-precise values, fast decision speed, scalability, etc. This article introduces a novel method for monitoring changes in the probabilistic distribution of multi-dimensional data streams. As a measure of the rapidity of changes, we analyze the popular Kullback-Leibler divergence. During the experimental study, we show how to use this metric to predict the concept drift occurrence and understand its nature. The obtained results encourage further work on the proposed methods and its application in the real tasks where the prediction of the future appearance of concept drift plays a crucial role, such as predictive maintenance.
A large body of recent work targets semantically conditioned image generation. Most such methods focus on the narrower task of pose transfer and ignore the more challenging task of subject transfer that consists in not only transferring the pose but also the appearance and background. In this work, we introduce SCAM (Semantic Cross Attention Modulation), a system that encodes rich and diverse information in each semantic region of the image (including foreground and background), thus achieving precise generation with emphasis on fine details. This is enabled by the Semantic Attention Transformer Encoder that extracts multiple latent vectors for each semantic region, and the corresponding generator that exploits these multiple latents by using semantic cross attention modulation. It is trained only using a reconstruction setup, while subject transfer is performed at test time. Our analysis shows that our proposed architecture is successful at encoding the diversity of appearance in each semantic region. Extensive experiments on the iDesigner and CelebAMask-HD datasets show that SCAM outperforms SEAN and SPADE; moreover, it sets the new state of the art on subject transfer.
Processing large indoor scenes is a challenging task, as scan registration and camera trajectory estimation methods accumulate errors across time. As a result, the quality of reconstructed scans is insufficient for some applications, such as visual-based localization and navigation, where the correct position of walls is crucial. For many indoor scenes, there exists an image of a technical floorplan that contains information about the geometry and main structural elements of the scene, such as walls, partitions, and doors. We argue that such a floorplan is a useful source of spatial information, which can guide a 3D model optimization. The standard RGB-D 3D reconstruction pipeline consists of a tracking module applied to an RGB-D sequence and a bundle adjustment (BA) module that takes the posed RGB-D sequence and corrects the camera poses to improve consistency. We propose a novel optimization algorithm expanding conventional BA that leverages the prior knowledge about the scene structure in the form of a floorplan. Our experiments on the Redwood dataset and our self-captured data demonstrate that utilizing floorplan improves accuracy of 3D reconstructions.
The past few years have witnessed a remarkable advance in deep learning for EEG-based sleep stage classification (SSC). However, the success of these models is attributed to possessing a massive amount of labeled data for training, limiting their applicability in real-world scenarios. In such scenarios, sleep labs can generate a massive amount of data, but labeling these data can be expensive and time-consuming. Recently, the self-supervised learning (SSL) paradigm has shined as one of the most successful techniques to overcome the scarcity of labeled data. In this paper, we evaluate the efficacy of SSL to boost the performance of existing SSC models in the few-labels regime. We conduct a thorough study on three SSC datasets, and we find that fine-tuning the pretrained SSC models with only 5% of labeled data can achieve competitive performance to the supervised training with full labels. Moreover, self-supervised pretraining helps SSC models to be more robust to data imbalance and domain shift problems. The code is publicly available at \url{https://github.com/emadeldeen24/eval_ssl_ssc}.
Satellite images constitute a highly valuable and abundant resource for many real world applications. However, the labeled data needed to train most machine learning models are scarce and difficult to obtain. In this context, the current work investigates a fully unsupervised methodology that, given a temporal sequence of satellite images, creates a partition of the ground according to its semantic properties and their evolution over time. The sequences of images are translated into a grid of multivariate time series of embedded tiles. The embedding and the partitional clustering of these sequences of tiles are constructed in two iterative steps: In the first step, the embedding is able to extract the information of the sequences of tiles based on a geographical neighborhood, and the tiles are grouped into clusters. In the second step, the embedding is refined by using the neighborhood defined by the clusters, and the final clustering of the sequences of tiles is obtained. We illustrate the methodology by conducting the semantic clustering of a sequence of 20 satellite images of the region of Navarra (Spain). The results show that the clustering of multivariate time series is robust and contains trustful spatio-temporal semantic information about the region under study. We unveil the close connection that exists between the geographic and embedded spaces, and find out that the semantic properties attributed to these kinds of embeddings are fully exploited and even enhanced by the proposed clustering of time series.