Despite being designed for performance rather than cognitive plausibility, transformer language models have been found to be better at predicting metrics used to assess human language comprehension than language models with other architectures, such as recurrent neural networks. Based on how well they predict the N400, a neural signal associated with processing difficulty, we propose and provide evidence for one possible explanation - their predictions are affected by the preceding context in a way analogous to the effect of semantic facilitation in humans.

DrQ-v2 builds on DrQ, an off-policy actor-critic approach that uses data augmentation to learn directly from pixels. We introduce several improvements that yield state-of-the-art results on the DeepMind Control Suite. Notably, DrQ-v2 is able to solve complex humanoid locomotion tasks directly from pixel observations, previously unattained by model-free RL. DrQ-v2 is conceptually simple, easy to implement, and provides significantly better computational footprint compared to prior work, with the majority of tasks taking just 8 hours to train on a single GPU. Finally, we publicly release DrQ-v2's implementation to provide RL practitioners with a strong and computationally efficient baseline.

Detecting customized moments and highlights from videos given natural language (NL) user queries is an important but under-studied topic. One of the challenges in pursuing this direction is the lack of annotated data. To address this issue, we present the Query-based Video Highlights (QVHighlights) dataset. It consists of over 10,000 YouTube videos, covering a wide range of topics, from everyday activities and travel in lifestyle vlog videos to social and political activities in news videos. Each video in the dataset is annotated with: (1) a human-written free-form NL query, (2) relevant moments in the video w.r.t. the query, and (3) five-point scale saliency scores for all query-relevant clips. This comprehensive annotation enables us to develop and evaluate systems that detect relevant moments as well as salient highlights for diverse, flexible user queries. We also present a strong baseline for this task, Moment-DETR, a transformer encoder-decoder model that views moment retrieval as a direct set prediction problem, taking extracted video and query representations as inputs and predicting moment coordinates and saliency scores end-to-end. While our model does not utilize any human prior, we show that it performs competitively when compared to well-engineered architectures. With weakly supervised pretraining using ASR captions, Moment-DETR substantially outperforms previous methods. Lastly, we present several ablations and visualizations of Moment-DETR. Data and code is publicly available at https://github.com/jayleicn/moment_detr

Computational intelligence-based ocean characteristics forecasting applications, such as Significant Wave Height (SWH) prediction, are crucial for avoiding social and economic loss in coastal cities. Compared to the traditional empirical-based or numerical-based forecasting models, "soft computing" approaches, including machine learning and deep learning models, have shown numerous success in recent years. In this paper, we focus on enabling the deep learning model to learn both short-term and long-term spatial-temporal dependencies for SWH prediction. A Wavelet Graph Neural Network (WGNN) approach is proposed to integrate the advantages of wavelet transform and graph neural network. Several parallel graph neural networks are separately trained on wavelet decomposed data, and the reconstruction of each model's prediction forms the final SWH prediction. Experimental results show that the proposed WGNN approach outperforms other models, including the numerical models, the machine learning models, and several deep learning models.

The success of deep neural networks in real-world problems has prompted many attempts to explain their training dynamics and generalization performance, but more guiding principles for the training of neural networks are still needed. Motivated by the edge of chaos principle behind the optimal performance of neural networks, we study the role of various hyperparameters in modern neural network training algorithms in terms of the order-chaos phase diagram. In particular, we study a fully analytical feedforward neural network trained on the widely adopted Fashion-MNIST dataset, and study the dynamics associated with the hyperparameters in back-propagation during the training process. We find that for the basic algorithm of stochastic gradient descent with momentum, in the range around the commonly used hyperparameter values, clear scaling relations are present with respect to the training time during the ordered phase in the phase diagram, and the model's optimal generalization power at the edge of chaos is similar across different training parameter combinations. In the chaotic phase, the same scaling no longer exists. The scaling allows us to choose the training parameters to achieve faster training without sacrificing performance. In addition, we find that the commonly used model regularization method - weight decay - effectively pushes the model towards the ordered phase to achieve better performance. Leveraging on this fact and the scaling relations in the other hyperparameters, we derived a principled guideline for hyperparameter determination, such that the model can achieve optimal performance by saturating it at the edge of chaos. Demonstrated on this simple neural network model and training algorithm, our work improves the understanding of neural network training dynamics, and can potentially be extended to guiding principles of more complex model architectures and algorithms.

Effective high-dimensional representation learning necessitates properly exploiting the geometry of data [Bronstein et al., 2021]--otherwise, it is a cursed estimation problem. Indeed, early success stories of deep learning relied on imposing strong geometric assumptions, primarily that the data lives on a grid domain; either spatial or temporal. In these two respective settings, convolutional neural networks (CNNs) [LeCun et al., 1998] and recurrent neural networks (RNNs) [Hochreiter and Schmidhuber, 1997] have traditionally dominated. While both CNNs and RNNs are demonstrably powerful models, with many applications of high interest, it can be recognised that most data coming from nature cannot be natively represented on a grid. Recent years are marked with a gradual shift of attention towards models that admit a more generic class of geometric structures [Masci et al., 2015, Veličković et al., 2017, Cohen et al., 2018, Battaglia et al., 2018, de Haan et al., 2020, Satorras et al., 2021].

We focus on building robustness in the convolutions of neural visual classifiers, especially against natural perturbations like elastic deformations, occlusions and Gaussian noise. Existing CNNs show outstanding performance on clean images, but fail to tackle naturally occurring perturbations. In this paper, we start from elastic perturbations, which approximate (local) view-point changes of the object. We present elastically-augmented convolutions (EAConv) by parameterizing filters as a combination of fixed elastically-perturbed bases functions and trainable weights for the purpose of integrating unseen viewpoints in the CNN. We show on CIFAR-10 and STL-10 datasets that the general robustness of our method on unseen occlusion and Gaussian perturbations improves, while even improving the performance on clean images slightly without performing any data augmentation.

While face recognition has drawn much attention, a large number of algorithms and models have been proposed with applications to daily life, such as authentication for mobile payments, etc. Recently, deep learning methods have dominated in the field of face recognition with advantages in comparisons to conventional approaches and even the human perception. Despite the popular adoption of deep learning-based methods to the field, researchers and engineers frequently need to reproduce existing algorithms with unified implementations (i.e., the identical deep learning framework with standard implementations of operators and trainers) and compare the performance of face recognition methods under fair settings (i.e., the same set of evaluation metrics and preparation of datasets with tricks on/off), so as to ensure the reproducibility of experiments. To the end, we develop face.evoLVe

We present Megaverse, a new 3D simulation platform for reinforcement learning and embodied AI research. The efficient design of our engine enables physics-based simulation with high-dimensional egocentric observations at more than 1,000,000 actions per second on a single 8-GPU node. Megaverse is up to 70x faster than DeepMind Lab in fully-shaded 3D scenes with interactive objects. We achieve this high simulation performance by leveraging batched simulation, thereby taking full advantage of the massive parallelism of modern GPUs. We use Megaverse to build a new benchmark that consists of several single-agent and multi-agent tasks covering a variety of cognitive challenges. We evaluate model-free RL on this benchmark to provide baselines and facilitate future research. The source code is available at https://www.megaverse.info

In the process of intelligently segmenting foods in images using deep neural networks for diet management, data collection and labeling for network training are very important but labor-intensive tasks. In order to solve the difficulties of data collection and annotations, this paper proposes a food segmentation method applicable to real-world through synthetic data. To perform food segmentation on healthcare robot systems, such as meal assistance robot arm, we generate synthetic data using the open-source 3D graphics software Blender placing multiple objects on meal plate and train Mask R-CNN for instance segmentation. Also, we build a data collection system and verify our segmentation model on real-world food data. As a result, on our real-world dataset, the model trained only synthetic data is available to segment food instances that are not trained with 52.2% mask AP@all, and improve performance by +6.4%p after fine-tuning comparing to the model trained from scratch. In addition, we also confirm the possibility and performance improvement on the public dataset for fair analysis. Our code and pre-trained weights are avaliable online at: https://github.com/gist-ailab/Food-Instance-Segmentation