Autism spectrum disorder (ASD) is a congenital neurodevelopmental disorder caused by a disorder of the nervous system. ASD is mainly characterized by developmental disorders, as well as abnormalities in social skills, communication skills, interests, and behavioral patterns. ASD is a congenital developmental disorder that cannot be completely cured by existing medical methods, and only after-day intervention can relieve its symptoms. In addition, the sooner you get the intervention, the better the patient's recovery. However, when most patients with ASD show the symptoms, they have missed the best intervention period. In order to allow the subject to be diagnosed earlier with ASD, we applied deep learning techniques to evaluate neural networks in ASD under functional magnetic resonance imaging (fMRI) imaging. At the same time, we also compared the accuracy of detection under the principle of structural magnetic resonance imaging (sMRI) imaging. Our experiments show that neural networks perform much better on fMRI data than sMRI data.

Using a large training dataset to train a big and powerful model -- a typical practice in modern deep learning, often suffers from two major problems: the expensive and slow training process and the error-prone labels. The existing approaches, targeting either speeding up the training by selecting a subset of representative training instances (subset selection) or eliminating the negative effect of mislabels during training (mislabel detection), do not perform well in this scenario due to overlooking one of these two problems. To fill this gap, we propose Deem, a novel data-efficient framework that selects a subset of representative training instances under label uncertainty. The key idea is to leverage the metadata produced during deep learning training, e.g., training losses and gradients, to estimate the label uncertainty and select the representative instances. In particular, we model the problem of subset selection under uncertainty as a problem of finding a subset that closely approximates the gradient of the whole training data set derived on soft labels. We show that it is an NP-hard problem with submodular property and propose a low complexity algorithm to solve this problem with an approximate ratio. Training on this small subset thus improves the training efficiency while guaranteeing the model's accuracy. Moreover, we propose an efficient strategy to dynamically refine this subset during the iterative training process. Extensive experiments on 6 datasets and 10 baselines demonstrate that Deem accelerates the training process up to 10X without sacrificing the model accuracy.

Non-relational tables, such as spreadsheets and web tables, are widely used to store structured data, yet they are primarily designed for human reading rather than machine-driven analysis. These tables often contain hierarchical headers, merged cells, and implicit spatial semantics, making it difficult for large language models (LLMs) to correctly interpret their structure. Moreover, real-world non-relational tables are often large, leading to prohibitive token costs and poor scalability when processed directly by LLMs. To address these challenges, we present NOAH, a system for accurate and scalable analysis of large non-relational tables using natural language queries. Rather than loading the entire table into an LLM or transforming it into a relational format, NOAH constructs a Hierarchical Semantic Index (HSI) in an offline phase that explicitly captures the structure of row and column headers. HSI enables query-driven attribute linking, allowing NOAH to retrieve only the data regions relevant to a given query. Building on HSI, NOAH employs a context-aware data analysis pipeline that progressively resolves complex analytical queries through dynamic operator selection. To further improve efficiency, NOAH incorporates embedding-based hierarchy search, together with a recall-guaranteed dynamic thresholding strategy. We evaluate NOAH on real-world benchmarks for non-relational table analysis. Experimental results demonstrate that NOAH outperforms the best-performing baseline by 31.2% in accuracy, while scaling effectively to large tables and reducing LLM token consumption by at least 3.3×.