Text recognition methods are gaining rapid development. Some advanced techniques, e.g., powerful modules, language models, and un- and semi-supervised learning schemes, consecutively push the performance on public benchmarks forward. However, the problem of how to better optimize a text recognition model from the perspective of loss functions is largely overlooked. CTC-based methods, widely used in practice due to their good balance between performance and inference speed, still grapple with accuracy degradation. This is because CTC loss emphasizes the optimization of the entire sequence target while neglecting to learn individual characters. We propose a self-distillation scheme for CTC-based model to address this issue. It incorporates a framewise regularization term in CTC loss to emphasize individual supervision, and leverages the maximizing-a-posteriori of latent alignment to solve the inconsistency problem that arises in distillation between CTC-based models. We refer to the regularized CTC loss as Distillation Connectionist Temporal Classification (DCTC) loss. DCTC loss is module-free, requiring no extra parameters, longer inference lag, or additional training data or phases. Extensive experiments on public benchmarks demonstrate that DCTC can boost text recognition model accuracy by up to 2.6%, without any of these drawbacks.
Automatic bill payment is an important part of business operations in fintech companies. The practice of deduction was mainly based on the total amount or heuristic search by dividing the bill into smaller parts to deduct as much as possible. This article proposes an end-to-end approach of automatically learning the optimal deduction paths (deduction amount in order), which reduces the cost of manual path design and maximizes the amount of successful deduction. Specifically, in view of the large search space of the paths and the extreme sparsity of historical successful deduction records, we propose a deep hierarchical reinforcement learning approach which abstracts the action into a two-level hierarchical space: an upper agent that determines the number of steps of deductions each day and a lower agent that decides the amount of deduction at each step. In such a way, the action space is structured via prior knowledge and the exploration space is reduced. Moreover, the inherited information incompleteness of the business makes the environment just partially observable. To be precise, the deducted amounts indicate merely the lower bounds of the available account balance. To this end, we formulate the problem as a partially observable Markov decision problem (POMDP) and employ an environment correction algorithm based on the characteristics of the business. In the world's largest electronic payment business, we have verified the effectiveness of this scheme offline and deployed it online to serve millions of users.
We present SDTracker, a method that harnesses the potential of synthetic data for multi-object tracking of real-world scenes in a domain generalization and semi-supervised fashion. First, we use the ImageNet dataset as an auxiliary to randomize the style of synthetic data. With out-of-domain data, we further enforce pyramid consistency loss across different "stylized" images from the same sample to learn domain invariant features. Second, we adopt the pseudo-labeling method to effectively utilize the unlabeled MOT17 training data. To obtain high-quality pseudo-labels, we apply proximal policy optimization (PPO2) algorithm to search confidence thresholds for each sequence. When using the unlabeled MOT17 training set, combined with the pure-motion tracking strategy upgraded via developed post-processing, we finally reach 61.4 HOTA.
We propose a novel framework to reconstruct accurate appearance and geometry with neural radiance fields (NeRF) for interacting hands, enabling the rendering of photo-realistic images and videos for gesture animation from arbitrary views. Given multi-view images of a single hand or interacting hands, an off-the-shelf skeleton estimator is first employed to parameterize the hand poses. Then we design a pose-driven deformation field to establish correspondence from those different poses to a shared canonical space, where a pose-disentangled NeRF for one hand is optimized. Such unified modeling efficiently complements the geometry and texture cues in rarely-observed areas for both hands. Meanwhile, we further leverage the pose priors to generate pseudo depth maps as guidance for occlusion-aware density learning. Moreover, a neural feature distillation method is proposed to achieve cross-domain alignment for color optimization. We conduct extensive experiments to verify the merits of our proposed HandNeRF and report a series of state-of-the-art results both qualitatively and quantitatively on the large-scale InterHand2.6M dataset.
Fault detection and diagnosis is significant for reducing maintenance costs and improving health and safety in chemical processes. Convolution neural network (CNN) is a popular deep learning algorithm with many successful applications in chemical fault detection and diagnosis tasks. However, convolution layers in CNN are very sensitive to the order of features, which can lead to instability in the processing of tabular data. Optimal order of features result in better performance of CNN models but it is expensive to seek such optimal order. In addition, because of the encapsulation mechanism of feature extraction, most CNN models are opaque and have poor interpretability, thus failing to identify root-cause features without human supervision. These difficulties inevitably limit the performance and credibility of CNN methods. In this paper, we propose an order-invariant and interpretable hierarchical dilated convolution neural network (HDLCNN), which is composed by feature clustering, dilated convolution and the shapley additive explanations (SHAP) method. The novelty of HDLCNN lies in its capability of processing tabular data with features of arbitrary order without seeking the optimal order, due to the ability to agglomerate correlated features of feature clustering and the large receptive field of dilated convolution. Then, the proposed method provides interpretability by including the SHAP values to quantify feature contribution. Therefore, the root-cause features can be identified as the features with the highest contribution. Computational experiments are conducted on the Tennessee Eastman chemical process benchmark dataset. Compared with the other methods, the proposed HDLCNN-SHAP method achieves better performance on processing tabular data with features of arbitrary order, detecting faults, and identifying the root-cause features.
Humans apprehend the world through various sensory modalities, yet language is their predominant communication channel. Machine learning systems need to draw on the same multimodal richness to have informed discourses with humans in natural language; this is particularly true for systems specialized in visually-dense information, such as dialogue, recommendation, and search engines for clothing. To this end, we train a visual question answering (VQA) system to answer complex natural language questions about apparel in fashion photoshoot images. The key to the successful training of our VQA model is the automatic creation of a visual question-answering dataset with 168 million samples from item attributes of 207 thousand images using diverse templates. The sample generation employs a strategy that considers the difficulty of the question-answer pairs to emphasize challenging concepts. Contrary to the recent trends in using several datasets for pretraining the visual question answering models, we focused on keeping the dataset fixed while training various models from scratch to isolate the improvements from model architecture changes. We see that using the same transformer for encoding the question and decoding the answer, as in language models, achieves maximum accuracy, showing that visual language models (VLMs) make the best visual question answering systems for our dataset. The accuracy of the best model surpasses the human expert level, even when answering human-generated questions that are not confined to the template formats. Our approach for generating a large-scale multimodal domain-specific dataset provides a path for training specialized models capable of communicating in natural language. The training of such domain-expert models, e.g., our fashion VLM model, cannot rely solely on the large-scale general-purpose datasets collected from the web.
Data-driven predictive methods which can efficiently and accurately transform protein sequences into biologically active structures are highly valuable for scientific research and therapeutical development. Determining accurate folding landscape using co-evolutionary information is fundamental to the success of modern protein structure prediction methods. As the state of the art, AlphaFold2 has dramatically raised the accuracy without performing explicit co-evolutionary analysis. Nevertheless, its performance still shows strong dependence on available sequence homologs. We investigated the cause of such dependence and presented EvoGen, a meta generative model, to remedy the underperformance of AlphaFold2 for poor MSA targets. EvoGen allows us to manipulate the folding landscape either by denoising the searched MSA or by generating virtual MSA, and helps AlphaFold2 fold accurately in low-data regime or even achieve encouraging performance with single-sequence predictions. Being able to make accurate predictions with few-shot MSA not only generalizes AlphaFold2 better for orphan sequences, but also democratizes its use for high-throughput applications. Besides, EvoGen combined with AlphaFold2 yields a probabilistic structure generation method which could explore alternative conformations of protein sequences, and the task-aware differentiable algorithm for sequence generation will benefit other related tasks including protein design.
Proteins are essential component of human life and their structures are important for function and mechanism analysis. Recent work has shown the potential of AI-driven methods for protein structure prediction. However, the development of new models is restricted by the lack of dataset and benchmark training procedure. To the best of our knowledge, the existing open source datasets are far less to satisfy the needs of modern protein sequence-structure related research. To solve this problem, we present the first million-level protein structure prediction dataset with high coverage and diversity, named as PSP. This dataset consists of 570k true structure sequences (10TB) and 745k complementary distillation sequences (15TB). We provide in addition the benchmark training procedure for SOTA protein structure prediction model on this dataset. We validate the utility of this dataset for training by participating CAMEO contest in which our model won the first place. We hope our PSP dataset together with the training benchmark can enable a broader community of AI/biology researchers for AI-driven protein related research.
Accurate and unbiased examinations of skin lesions are critical for the early diagnosis and treatment of skin conditions and disorders. Visual features of skin lesions vary significantly because the skin images are collected from patients with different skin colours and morphologies by using dissimilar imaging equipment. Recent studies have reported ensembled convolutional neural networks (CNNs) to classify the images for early diagnosis of skin disorders. However, the practical use of these ensembled CNNs is limited because they are heavyweight and inadequate for using contextual information. Although lightweight networks (e.g., MobileNetV3 and EfficientNet) were developed to achieve parameters reduction for implementing deep neural networks on mobile devices, insufficient depth of feature representation restricts the performance. To address the existing limitations, we introduce a new lite and effective neural network, namely HierAttn. The HierAttn applies a novel strategy to learn the local and global features by using multi-stage and multi-branch attention mechanisms. The efficacy of HierAttn was evaluated by using the dermoscopy images dataset ISIC2019 and smartphone photos dataset PAD-UFES-20. The experimental results show that HierAttn achieves the best top-1 accuracy and AUC among the state-of-the-art lightweight networks. The code is available at https://github.com/anthonyweidai/HierAttn.