Deep learning techniques have been successfully applied in handwriting recognition. Oracle bone inscriptions (OBI) are the earliest hieroglyphs in China and valuable resources for studying the etymology of Chinese characters. OBI are of important historical and cultural value in China; thus, textual research surrounding the characters of OBI is a huge challenge for archaeologists. In this work, we built a dataset named OBI-100, which contains 100 classes of oracle bone inscriptions collected from two OBI dictionaries. The dataset includes more than 128,000 character samples related to the natural environment, humans, animals, plants, etc. In addition, we propose improved models based on three typical deep convolutional network structures to recognize the OBI-100 dataset. By modifying the parameters, adjusting the network structures, and adopting optimization strategies, we demonstrate experimentally that these models perform fairly well in OBI recognition. For the 100-category OBI classification task, the optimal model achieves an accuracy of 99.5%, which shows competitive performance compared with other state-of-the-art approaches. We hope that this work can provide a valuable tool for character recognition of OBI.

In this paper, we propose a safe reinforcement learning approach to synthesize deep neural network (DNN) controllers for nonlinear systems subject to safety constraints. The proposed approach employs an iterative scheme where alearnerand averifierinteract to synthesize safe DNN controllers. Thelearnertrains a DNN controller via deep reinforcement learning, and theverifiercertifies the learned controller through computing a maximal safe initial region and its corresponding barrier certificate, based on polynomial abstraction and bilinear matrix inequalities solving. Compared with the existing verification-in-the-loop synthesis methods, our iterative framework is a sequential synthesis scheme of controllers and barrier certificates, which can learn safe controllers with adaptive barrier certificates rather than user-defined ones. We implement the tool SRLBC and evaluate its performance over a set of benchmark examples. The experimental results demonstrate that our approach efficiently synthesizes safe DNN controllers even for a nonlinear system with dimension up to 12.

Automatic geometric problem-solving is an active and challenging subfield at the intersection of AI and mathematics, where geometric problem parsing plays a critical role. It involves converting geometric diagram and text into certain formal language. Due to the complexity of geometric shapes and the diversity of geometric relationships, geometric problem parsing demands that the parser exhibit cross-modal comprehension and reasoning capabilities. In this paper, we propose an enhanced geometric problem parsing method called FGeo-Parser, which converts problem diagrams and text into the formal language of the FormalGeo. It also supports reverse formalization to generate human-like solutions, reflecting the symmetry between parsing and generating. Specifically, diagram parser leverages the BLIP to generate the construction CDL and image CDL, while text parser employs the T5 to produce the text CDL and goal CDL where these neural networks are both based on a symmetric encoder–decoder architecture. With the assistance of a theorem predictor, these CDLs were automatically parsed and step-by-step reasoning was executed within FGPS. Finally, the reasoning process was input into a solution generator, which subsequently produced a human-like solution process. Additionally, we re-annotated problem diagrams and text based on the FormalGeo7K dataset. The formalization experiments on the new dataset achieved a match accuracy of 91.51% and a perfect accuracy of 56.47%, while the combination with the theorem predictor achieved a problem-solving accuracy of 63.45%.

The global outbreak of the COVID-19 epidemic has become a major public health problem. COVID-19 virus infection triggers a complex immune response. CD8⁺ T cells, in particular, play an essential role in controlling the severity of the disease. However, the mechanism of the regulatory role of CD8⁺ T cells on COVID-19 remains poorly investigated. In this study, single-cell gene expression profiles from three CD8⁺ T cell subtypes (effector, memory, and naive T cells) were downloaded. Each cell subtype included three disease states, namely, acute COVID-19, convalescent COVID-19, and unexposed individuals. The profiles on each cell subtype were individually analyzed in the same way. Irrelevant features in the profiles were first excluded by the Boruta method. The remaining features for each CD8⁺ T cells subtype were further analyzed by Max-Relevance and Min-Redundancy, Monte Carlo feature selection, and light gradient boosting machine methods to obtain three feature lists. These lists were then brought into the incremental feature selection method to determine the optimal features for each cell subtype. Their corresponding genes may be latent biomarkers to determine COVID-19 severity. Genes, such as ZFP36, DUSP1, TCR, and IL7R, can be confirmed to play an immune regulatory role in COVID-19 infection and recovery. The results of functional enrichment analysis revealed that these important genes may be associated with immune functions, such as response to cAMP, response to virus, T cell receptor complex, T cell activation, and T cell differentiation. This study further set up different gene expression pattens, represented by classification rules, on three states of COVID-19 and constructed several efficient classifiers to distinguish COVID-19 severity. The findings of this study provided new insights into the biological processes of CD8⁺ T cells in regulating the immune response.

With the increasing number of people suffering from cancer, this illness has become a major health problem worldwide. Exploring the biological functions and signaling pathways of carcinogenesis is essential for cancer detection and research. In this study, a mutation dataset for eleven cancer types was first obtained from a web-based resource called cBioPortal for Cancer Genomics, followed by extracting 21,049 features from three aspects: relationship to GO and KEGG (enrichment features), mutated genes learned by word2vec (text features), and protein-protein interaction network analyzed by node2vec (network features). Irrelevant features were then excluded using the Boruta feature filtering method, and the retained relevant features were ranked by four feature selection methods (least absolute shrinkage and selection operator, minimum redundancy maximum relevance, Monte Carlo feature selection and light gradient boosting machine) to generate four feature-ranked lists. Incremental feature selection was used to determine the optimal number of features based on these feature lists to build the optimal classifiers and derive interpretable classification rules. The results of four feature-ranking methods were integrated to identify key functional pathways, such as olfactory transduction (hsa04740) and colorectal cancer (hsa05210), and the roles of these functional pathways in cancers were discussed in reference to literature. Overall, this machine learning-based study revealed the altered biological functions of cancers and provided a reference for the mechanisms of different cancers.

Deep convolutional neural networks (CNNs) have achieved unprecedented success in single image super-resolution over the past few years. Meanwhile, there is an increasing demand for single image super-resolution with arbitrary scale factors in real-world scenarios. Many approaches adopt scale-specific multi-path learning to cope with multi-scale super-resolution with a single network. However, these methods require a large number of parameters. To achieve a better balance between the reconstruction quality and parameter amounts, we proposes a learnable interpolation method that leverages the advantages of neural networks and interpolation methods to tackle the scale-arbitrary super-resolution task. The scale factor is treated as a function parameter for generating the kernel weights for the learnable interpolation. We demonstrate that the learnable interpolation builds a bridge between neural networks and traditional interpolation methods. Experiments show that the proposed learnable interpolation requires much fewer parameters and outperforms state-of-the-art super-resolution methods.