We investigate the use of models from the theory of regularity structure as features in machine learning tasks. A model is a multi-linear function of a space-time signal designed to well-approximate solutions to partial differential equations (PDEs), even in low regularity regimes. Models can be seen as natural multi-dimensional generalisations of signatures of paths; our work therefore aims to extend the recent use of signatures in data science beyond the context of time-ordered data. We provide a flexible definition of a model feature vector associated to a space-time signal, along with two algorithms which illustrate ways in which these features can be combined with linear regression. We apply these algorithms in several numerical experiments designed to learn solutions to PDEs with a given forcing and boundary data. Our experiments include semi-linear parabolic and wave equations with forcing, and Burgers' equation with no forcing. We find an advantage in favour of our algorithms when compared to several alternative methods. Additionally, in the experiment with Burgers' equation, we noticed stability in the prediction power when noise is added to the observations.

How do the neural networks distinguish two images? It is of critical importance to understand the matching mechanism of deep models for developing reliable intelligent systems for many risky visual applications such as surveillance and access control. However, most existing deep metric learning methods match the images by comparing feature vectors, which ignores the spatial structure of images and thus lacks interpretability. In this paper, we present a deep interpretable metric learning (DIML) method for more transparent embedding learning. Unlike conventional metric learning methods based on feature vector comparison, we propose a structural matching strategy that explicitly aligns the spatial embeddings by computing an optimal matching flow between feature maps of the two images. Our method enables deep models to learn metrics in a more human-friendly way, where the similarity of two images can be decomposed to several part-wise similarities and their contributions to the overall similarity. Our method is model-agnostic, which can be applied to off-the-shelf backbone networks and metric learning methods. We evaluate our method on three major benchmarks of deep metric learning including CUB200-2011, Cars196, and Stanford Online Products, and achieve substantial improvements over popular metric learning methods with better interpretability. Code is available at https://github.com/wl-zhao/DIML

Large-scale pretraining of visual representations has led to state-of-the-art performance on a range of benchmark computer vision tasks, yet the benefits of these techniques at extreme scale in complex production systems has been relatively unexplored. We consider the case of a popular visual discovery product, where these representations are trained with multi-task learning, from use-case specific visual understanding (e.g. skin tone classification) to general representation learning for all visual content (e.g. embeddings for retrieval). In this work, we describe how we (1) generate a dataset with over a billion images via large weakly-supervised pretraining to improve the performance of these visual representations, and (2) leverage Transformers to replace the traditional convolutional backbone, with insights into both system and performance improvements, especially at 1B+ image scale. To support this backbone model, we detail a systematic approach to deriving weakly-supervised image annotations from heterogenous text signals, demonstrating the benefits of clustering techniques to handle the long-tail distribution of image labels. Through a comprehensive study of offline and online evaluation, we show that large-scale Transformer-based pretraining provides significant benefits to industry computer vision applications. The model is deployed in a production visual shopping system, with 36% improvement in top-1 relevance and 23% improvement in click-through volume. We conduct extensive experiments to better understand the empirical relationships between Transformer-based architectures, dataset scale, and the performance of production vision systems.

Over-parameterized deep networks trained using gradient-based optimizers are a popular choice for solving classification and ranking problems. Without appropriately tuned $\ell_2$ regularization or weight decay, such networks have the tendency to make output scores (logits) and network weights large, causing training loss to become too small and the network to lose its adaptivity (ability to move around) in the parameter space. Although regularization is typically understood from an overfitting perspective, we highlight its role in making the network more adaptive and enabling it to escape more easily from weights that generalize poorly. To provide such a capability, we propose a method called Logit Attenuating Weight Normalization (LAWN), that can be stacked onto any gradient-based optimizer. LAWN controls the logits by constraining the weight norms of layers in the final homogeneous sub-network. Empirically, we show that the resulting LAWN variant of the optimizer makes a deep network more adaptive to finding minimas with superior generalization performance on large-scale image classification and recommender systems. While LAWN is particularly impressive in improving Adam, it greatly improves all optimizers when used with large batch sizes

Deep learning is gaining instant popularity in computer aided diagnosis of COVID-19. Due to the high sensitivity of Computed Tomography (CT) to this disease, CT-based COVID-19 detection with visual models is currently at the forefront of medical imaging research. Outcomes published in this direction are frequently claiming highly accurate detection under deep transfer learning. This is leading medical technologists to believe that deep transfer learning is the mainstream solution for the problem. However, our critical analysis of the literature reveals an alarming performance disparity between different published results. Hence, we conduct a systematic thorough investigation to analyze the effectiveness of deep transfer learning for COVID-19 detection with CT images. Exploring 14 state-of-the-art visual models with over 200 model training sessions, we conclusively establish that the published literature is frequently overestimating transfer learning performance for the problem, even in the prestigious scientific sources. The roots of overestimation trace back to inappropriate data curation. We also provide case studies that consider more realistic scenarios, and establish transparent baselines for the problem. We hope that our reproducible investigation will help in curbing hype-driven claims for the critical problem of COVID-19 diagnosis, and pave the way for a more transparent performance evaluation of techniques for CT-based COVID-19 detection.

The purpose of the Session-Based Recommendation System is to predict the user's next click according to the previous session sequence. The current studies generally learn user preferences according to the transitions of items in the user's session sequence. However, other effective information in the session sequence, such as user profiles, are largely ignored which may lead to the model unable to learn the user's specific preferences. In this paper, we propose a heterogeneous graph neural network-based session recommendation method, named SR-HetGNN, which can learn session embeddings by heterogeneous graph neural network (HetGNN), and capture the specific preferences of anonymous users. Specifically, SR-HetGNN first constructs heterogeneous graphs containing various types of nodes according to the session sequence, which can capture the dependencies among items, users, and sessions. Second, HetGNN captures the complex transitions between items and learns the item embeddings containing user information. Finally, to consider the influence of users' long and short-term preferences, local and global session embeddings are combined with the attentional network to obtain the final session embedding. SR-HetGNN is shown to be superior to the existing state-of-the-art session-based recommendation methods through extensive experiments over two real large datasets Diginetica and Tmall.

The increased memory and processing capabilities of today's edge devices create opportunities for greater edge intelligence. In the domain of vision, the ability to adapt a Convolutional Neural Network's (CNN) structure and parameters to the input data distribution leads to systems with lower memory footprint, latency and power consumption. However, due to the limited compute resources and memory budget on edge devices, it is necessary for the system to be able to predict the latency and memory footprint of the training process in order to identify favourable training configurations of the network topology and device combination for efficient network adaptation. This work proposes perf4sight, an automated methodology for developing accurate models that predict CNN training memory footprint and latency given a target device and network. This enables rapid identification of network topologies that can be retrained on the edge device with low resource consumption. With PyTorch as the framework and NVIDIA Jetson TX2 as the target device, the developed models predict training memory footprint and latency with 95% and 91% accuracy respectively for a wide range of networks, opening the path towards efficient network adaptation on edge GPUs.

Recommender systems aim to provide personalized services to users and are playing an increasingly important role in our daily lives. The key of recommender systems is to predict how likely users will interact with items based on their historical online behaviors, e.g., clicks, add-to-cart, purchases, etc. To exploit these user-item interactions, there are increasing efforts on considering the user-item interactions as a user-item bipartite graph and then performing information propagation in the graph via Graph Neural Networks (GNNs). Given the power of GNNs in graph representation learning, these GNN-based recommendation methods have remarkably boosted the recommendation performance. Despite their success, most existing GNN-based recommender systems overlook the existence of interactions caused by unreliable behaviors (e.g., random/bait clicks) and uniformly treat all the interactions, which can lead to sub-optimal and unstable performance. In this paper, we investigate the drawbacks (e.g., non-adaptive propagation and non-robustness) of existing GNN-based recommendation methods. To address these drawbacks, we propose the Graph Trend Networks for recommendations (GTN) with principled designs that can capture the adaptive reliability of the interactions. Comprehensive experiments and ablation studies are presented to verify and understand the effectiveness of the proposed framework. Our implementation and datasets can be released after publication.

A number of deep learning based algorithms have been proposed to recover high-quality videos from low-quality compressed ones. Among them, some restore the missing details of each frame via exploring the spatiotemporal information of neighboring frames. However, these methods usually suffer from a narrow temporal scope, thus may miss some useful details from some frames outside the neighboring ones. In this paper, to boost artifact removal, on the one hand, we propose a Recursive Fusion (RF) module to model the temporal dependency within a long temporal range. Specifically, RF utilizes both the current reference frames and the preceding hidden state to conduct better spatiotemporal compensation. On the other hand, we design an efficient and effective Deformable Spatiotemporal Attention (DSTA) module such that the model can pay more effort on restoring the artifact-rich areas like the boundary area of a moving object. Extensive experiments show that our method outperforms the existing ones on the MFQE 2.0 dataset in terms of both fidelity and perceptual effect. Code is available at https://github.com/zhaominyiz/RFDA-PyTorch.

Recent advances in multi-fingered robotic grasping have enabled fast 6-Degrees-Of-Freedom (DOF) single object grasping. Multi-finger grasping in cluttered scenes, on the other hand, remains mostly unexplored due to the added difficulty of reasoning over obstacles which greatly increases the computational time to generate high-quality collision-free grasps. In this work we address such limitations by introducing DDGC, a fast generative multi-finger grasp sampling method that can generate high quality grasps in cluttered scenes from a single RGB-D image. DDGC is built as a network that encodes scene information to produce coarse-to-fine collision-free grasp poses and configurations. We experimentally benchmark DDGC against the simulated-annealing planner in GraspIt! on 1200 simulated cluttered scenes and 7 real world scenes. The results show that DDGC outperforms the baseline on synthesizing high-quality grasps and removing clutter while being 5 times faster. This, in turn, opens up the door for using multi-finger grasps in practical applications which has so far been limited due to the excessive computation time needed by other methods.