Yuxiang Wei

Alzheimer's disease (AD) progresses from asymptomatic changes to clinical symptoms, underscoring the critical need for early detection to facilitate timely treatment. Functional magnetic resonance imaging (fMRI) offers non-invasive biomarkers for detection, but current methods fail to reliably identify pre-symptomatic individuals due to two key challenges: (1) Subtle, anatomically distinct fMRI patterns in pre-symptomatic cases that resemble healthy controls more than symptomatic patients, and (2) Severe class imbalance in real-world data, where healthy controls vastly outnumber pre-symptomatic subjects. To address this, we reconceptualize AD diagnosis as a multi-stage distillation task, where insights from easier-to-detect symptomatic cases guide pre-symptomatic detection. We propose a novel margin-aware knowledge distillation (KD) framework with two innovations: (1) We leverage Unbalanced Optimal Transport (UOT) for Feature Distillation to flexibly adapt to anatomical differences in brain patterns caused by neurodegeneration and ensure effective distillation from later to earlier disease stages. (2) We propose Self-Distillation with Dynamic Margins to combat class imbalance by adaptively refining the classification boundary. We evaluate our proposed framework across four distinct base models and demonstrate its superiority over state-of-the-art KD methods. Additionally, we show the significance of various brain regions in identifying pre-symptomatic subjects, as well as how features are transferred during distillation. These contributions advance the development of more precise diagnostic tools and foster a deeper understanding of early disease manifestations, marking a significant stride towards more reliable and earlier AD diagnosis.

Visual Prompt Tuning (VPT) has emerged as a parameter-efficient fine-tuning paradigm for vision transformers, with conventional approaches utilizing dataset-level prompts that remain the same across all input instances. We observe that this strategy results in sub-optimal performance due to high variance in downstream datasets. To address this challenge, we propose Visual Instance-aware Prompt Tuning (ViaPT), which generates instance-aware prompts based on each individual input and fuses them with dataset-level prompts, leveraging Principal Component Analysis (PCA) to retain important prompting information. Moreover, we reveal that VPT-Deep and VPT-Shallow represent two corner cases based on a conceptual understanding, in which they fail to effectively capture instance-specific information, while random dimension reduction on prompts only yields performance between the two extremes. Instead, ViaPT overcomes these limitations by balancing dataset-level and instance-level knowledge, while reducing the amount of learnable parameters compared to VPT-Deep. Extensive experiments across 34 diverse datasets demonstrate that our method consistently outperforms state-of-the-art baselines, establishing a new paradigm for analyzing and optimizing visual prompts for vision transformers.

As a key component in boosting online user growth, uplift modeling aims to measure individual user responses (e.g., whether to play the game) to various treatments, such as gaming bonuses, thereby enhancing business outcomes. However, previous research typically considers a single-task, single-treatment setting, where only one treatment exists and the overall treatment effect is measured by a single type of user response. In this paper, we propose a Multi-Treatment Multi-Task (MTMT) uplift network to estimate treatment effects in a multi-task scenario. We identify the multi-treatment problem as a causal inference problem with a tiered response, comprising a base effect (from offering a treatment) and an incremental effect (from offering a specific type of treatment), where the base effect can be numerically much larger than the incremental effect. Specifically, MTMT separately encodes user features and treatments. The user feature encoder uses a multi-gate mixture of experts (MMOE) network to encode relevant user features, explicitly learning inter-task relations. The resultant embeddings are used to measure natural responses per task. Furthermore, we introduce a treatment-user feature interaction module to model correlations between each treatment and user feature. Consequently, we separately measure the base and incremental treatment effect for each task based on the produced treatment-aware representations. Experimental results based on an offline public dataset and an online proprietary dataset demonstrate the effectiveness of MTMT in single/multi-treatment and single/multi-task settings. Additionally, MTMT has been deployed in our gaming platform to improve user experience.

Despite the availability of electronic design automation (EDA) tools, the design of analog circuits is still a complex process that requires a high level of expertise in circuit parameterization and topological optimization. In recent years, researchers have increasingly applied artificial intelligence (AI) to assist circuit design while employing simulations to evaluate the circuit performance. This is time consuming due to the time-domain simulation processes and the lack of using existing knowledge. To address these issues, we propose a graph theory based machine learning framework for analog circuits design automation. Specifically, we employ graph theory to evaluate the circuit transfer function and thus to avoid time-domain simulations, as well as for AI to accommodate circuit topology variations in addition to parameter optimization. For this, we develop a graph-based coding method to represent the circuit topology and thereby a parameter-topology optimizing algorithm to evolve the circuit. Experimental results show that this method efficiently generates high-quality circuits and offers multiple choices of candidate EDA topologies for practical implementations.

Wafer defect pattern recognition (DPR) is an essential process in semiconductor manufacturing. This helps the manufacturers improve their fabrication process, reduce initial costs, and avoid further defects. However, mixed-type defects can be varied and complicated, which is difficult for traditional deep learning model to learn their features. To address this problem, we utilize dynamicity to allow the adaptive integration of information. Therefore, this article proposed a mixed-type wafer DPR framework based on multifaceted dynamic convolution (MFD). This framework can effectively detect decisive features and integrate helpful information from defect patterns. Specifically, the proposed MFD first designs a 1-D autoencoder to process the input and extends it to a 3-D to obtain trans-pixel dynamic perception. In addition, we apply a channel-amalgamating mechanism to obtain the trans-channel dynamic perception and further synthesize the information across the channel. We appraised our model based on a dataset with 38 defect patterns generated from real industrial manufacturing. The results showed that the proposed framework excelled the vanilla convolutional neural network (CNN) and state-of-the-art deep learning models. Further interpretable analysis based on visualization also indicated that this framework can adaptively procure the significant features from different defects.

Defect pattern recognition (DPR) of wafer maps can be essential as the accurate classification helps with the fabrication process improvement and thus avoiding further defects. During fabrication, various defect patterns may be mixed. In contrast to single-type DPR, mixed-type DPR can be much more complicated due to the varied spatial features, numerous types, uncertain number of defects, etc. To effectively recognize the mixed-type defects, we proposed a novel multi-scale information fusion transformer framework (MSF-Trans). Specifically, an MSF-Network is proposed to focus on the detailed features of wafer maps, which can also selectively enhance valuable information. Subsequently, the Transformer architecture is introduced, which used multi-head attention mechanism to encode the global context features of wafer maps, thereby modeling the internal relationship between wafer maps and defect patterns. MSF-Trans fully integrates the advantages of convolutional network and transformer in detailed feature learning and global feature learning. MSF-Trans is evaluated on a real dataset with 38 defect patterns. The results show that MSF-Trans has excellent defect recognition ability and is significantly better than the existing deep learning algorithms. Further interpretable analysis indicates that MSF-Trans can effectively recognize the defect pattern and enhance learning of valuable information, which facilitates the recognition of mixed-type defects.