Sharpness-aware minimization (SAM) is widely recognized for enhancing the generalization performance of deep neural networks. However, recent works have challenged the statement that flatness implies generalization, demonstrating that it is insufficient as the indicator of generalization. In this paper, we reveal an insightful phenomenon: among minima of similar sharpness, stochastic optimization algorithms tend to prefer those with lower nonuniformity. We define nonuniformity by both the magnitude and structure of the gradient noise, and show that it fundamentally differs from sharpness and plays a critical role in generalization. Specifically, we first theoretically prove that the expected generalization gap of models trained via stochastic optimization algorithm is positively correlated with nonuniformity (the magnitude of the gradient noise). Empirically, we show that nonuniformity exhibits a stronger correlation with generalization than sharpness, especially in Transformer models. Furthermore, we demonstrate that the nonuniformity (the structure of the gradient noise) more effectively guides the algorithm towards sparser solutions and exhibits better generalization performance than sharpness-based methods in the high-dimensional sparse regression problem. Finally, extensive experiments on various datasets and models confirm the advantages of nonuniformity for generalization: (1) optimization guided by nonuniformity achieves better generalization compared to those achieved through flatness (including standard training, transfer learning, hyperparameter sensitivity and robustness to label noise); (2) model architecture (such as depth and width) is closely related to nonuniformity.

Retrosynthesis, which predicts the reactants of a given target molecule, is an essential task for drug discovery. Retrosynthesis prediction based on molecular graph editing has garnered widespread attention due to excellent interpretability. Existing methods fail to effectively incorporate the chemical knowledge when learning molecular representations. To address this issue, we propose a Knowledge-enhanced Graph Contrastive Learning model (KGCL), which retrieve functional group embeddings from a chemical knowledge graph and integrate them into the atomic embeddings of the product molecule using an attention mechanism. Furthermore, we introduce a graph contrastive learning strategy that generates augmented samples using graph edits to improve the molecular graph encoder. Our proposed method outperforms the strong baseline method Graph2Edits by 1.6% and 3.2% in terms of the top-1 accuracy and top-1 round-trip accuracy on the USPTO-50K dataset, respectively, and also achieves a new state-of-the-art performance among semi-template-based methods on the USPTO-FULL dataset.

Teaching discourse can not only introduce the class into different teaching episodes, but also interact emotionally with students by different instructions. Therefore, the richness of teaching discourse is a very important evaluation criteria. This paper proposes a teaching discourse richness detection model based on prompt learning ProTDR, which converts the data of downstream tasks into natural language form by setting different kinds of prompt templates. Without changing the structure of the backbone model, the prompt learning can assist the pre-trained model to better judge the teaching discourse category. We conduct a series of experiments on a Chinese DialogId dataset which consists of teacher utterance transcripts in real classrooms. Compared with current mainstream pre-trained models, and the proposed model achieves the state-of-the-art performance on the DialogId dataset. Finally, we give a test sample of the model ProTDR on a real classroom teaching of a high school, and visually prove the feasibility and effectiveness of the proposed model.

In this study, a natural scene text detection method based on the improved faster region-based convolutional neural network (R-CNN) is proposed. This method extracts image features with the Inception-ResNet architecture, adopts a region proposal network to generate region proposals for the extracted features, merges the fine-tuned features with the region proposals, and finally, uses Fast R-CNN to classify and locate text. The proposed method solves the problems of varying text sizes and the text being obscured in the image. Compared with the original Faster R-CNN, the multilevel Inception-ResNet network model presented in this study can extract deeper text features. The extracted feature map is further sparsely represented by Reduction B, Inception ResNet C and Avg Pool, and then is fused with text regions obtained by the text feature mapping lower layer network to acquire the exact text regions. The text detection method presented in this study is tested on the 2017 dataset of ICDAR2017 Competition on Reading Chinese Text in the Wild (RCTW-17), which contains a large number of distorted, blurry, different scale and size texts. An accuracy of 76.4% is achieved in this platform, thereby proving the efficiency of the proposed method.

Hidden Markov model (HMM) is the dominant technology in speech recognition. The problem of optimizing model parameters is of great interest to the researchers in this area. The Baum-Welch (BW) algorithm is a popular estimation method due to its reliability and efficiency. However, it is easily trapped in local optimum. Recently, genetic algorithm (GA) and particle swarm optimization (PSO) have attracted considerable attention among various modern heuristic optimization techniques. Since the two approaches are supposed to find a solution to a given objective function but employ different strategies and computational effort, it is appropriate to compare their performance. This paper presents the application and performance comparison of PSO and GA for continuous HMM optimization in continuous speech recognition. The experimental results demonstrate that PSO is superior to GA in respect of the recognition performance.