Zechun Cao

Vision-language foundation models have emerged as powerful general-purpose representation learners with strong potential for multimodal understanding, but their deterministic embeddings often fail to provide the reliability required for high-stakes biomedical applications. This work introduces MedProbCLIP, a probabilistic vision-language learning framework for chest X-ray and radiology report representation learning and bidirectional retrieval. MedProbCLIP models image and text representations as Gaussian embeddings through a probabilistic contrastive objective that explicitly captures uncertainty and many-to-many correspondences between radiographs and clinical narratives. A variational information bottleneck mitigates overconfident predictions, while MedProbCLIP employs multi-view radiograph encoding and multi-section report encoding during training to provide fine-grained supervision for clinically aligned correspondence, yet requires only a single radiograph and a single report at inference. Evaluated on the MIMIC-CXR dataset, MedProbCLIP outperforms deterministic and probabilistic baselines, including CLIP, CXR-CLIP, and PCME++, in both retrieval and zero-shot classification. Beyond accuracy, MedProbCLIP demonstrates superior calibration, risk-coverage behavior, selective retrieval reliability, and robustness to clinically relevant corruptions, underscoring the value of probabilistic vision-language modeling for improving the trustworthiness and safety of radiology image-text retrieval systems.

Rust’s growing popularity in high-integrity systems requires automated vulnerability detection in order to maintain its strong safety guarantees. Although Rust’s ownership model and compile-time checks prevent many errors, sometimes unexpected bugs may occasionally pass analysis, underlining the necessity for automated safe and unsafe code detection. This paper presents Rust-IR-BERT, a machine learning approach to detect security vulnerabilities in Rust code by analyzing its compiled LLVM intermediate representation (IR) instead of the raw source code. This approach offers novelty by employing LLVM IR’s language-neutral, semantically rich representation of the program, facilitating robust detection by capturing core data and control-flow semantics and reducing language-specific syntactic noise. Our method leverages a graph-based transformer model, GraphCodeBERT, which is a transformer architecture pretrained model to encode structural code semantics via data-flow information, followed by a gradient boosting classifier, CatBoost, that is capable of handling complex feature interactions—to classify code as vulnerable or safe. The model was evaluated using a carefully curated dataset of over 2300 real-world Rust code samples (vulnerable and non-vulnerable Rust code snippets) from RustSec and OSV advisory databases, compiled to LLVM IR and labeled with corresponding Common Vulnerabilities and Exposures (CVEs) identifiers to ensure comprehensive and realistic coverage. Rust-IR-BERT achieved an overall accuracy of 98.11%, with a recall of 99.31% for safe code and 93.67% for vulnerable code. Despite these promising results, this study acknowledges potential limitations such as focusing primarily on known CVEs. Built on a representative dataset spanning over 2300 real-world Rust samples from diverse crates, Rust-IR-BERT delivers consistently strong performance. Looking ahead, practical deployment could take the form of a Cargo plugin or pre-commit hook that automatically generates and scans LLVM IR artifacts during the development cycle, enabling developers to catch vulnerabilities at an early stage in the development cycle.

The burgeoning sophistication of Artificial Intelligence (AI) has catalyzed the rapid proliferation of Large Language Models (LLMs) within software development. These models are increasingly employed to automate the generation of functionally correct code, address complex computational problems, and facilitate the debugging of existing software systems. However, LLM-generated code often faces challenges due to inherent inefficiencies, including redundant logical structures, factually inconsistent content (hallucinations), and programming errors. To address this issue, our research rigorously evaluated the computational efficiency of Python code generated by three prominent LLMs: GPT-4o-Mini, GPT-3.5-Turbo, and GPT-4-Turbo. The evaluation metrics encompass execution time, memory utilization, and peak memory consumption, while maintaining the functional correctness of the generated code. Leveraging the EffiBench benchmark datasets within the Google Vertex AI Workbench environment, across a spectrum of machine configurations, the study implemented a consistent seed parameter to ensure experimental reproducibility. Furthermore, we investigated the impact of two distinct optimization strategies: Chain-of-Thought (CoT) prompting and model fine-tuning. Our findings reveal a significant enhancement in efficiency metrics for GPT-4o-Mini and GPT-3.5-Turbo when employing CoT prompting; however, this trend was not observed for GPT-4-Turbo. Based on its promising performance with CoT prompting, we selected the GPT-4o-Mini model for subsequent fine-tuning, aiming to further enhance both its computational efficiency and accuracy. However, contrary to our expectations, fine-tuning the GPT-4o-Mini model led to a discernible degradation in both its accuracy and computational efficiency. In conclusion, this study provides empirical evidence suggesting that the deployment of high-CPU machine configurations, in synergy with the utilization of the GPT-4o-Mini model and CoT prompting techniques, yields demonstrably more efficient and accurate LLM-generated Python code, particularly within computationally intensive application scenarios.

Rust’s growing popularity in high-integrity systems requires automated vulnerability detection in order to maintain its strong safety guarantees. Although Rust’s ownership model and compile-time checks prevent many errors, sometimes unexpected bugs may occasionally pass analysis, underlining the necessity for automated safe and unsafe code detection. This paper presents Rust-IR-BERT, a machine learning approach to detect security vulnerabilities in Rust code by analyzing its compiled LLVM intermediate representation (IR) instead of the raw source code. Using LLVM IR provides a language-neutral, semantically rich view of the program, capturing data and control flow, and reducing the noise of high-level syntax differences. Our method leverages a transformer model, GraphCodeBERT, to embed the IR and CatBoost classifier to classify code as vulnerable or safe. When evaluated on a mix of known buggy and safe code, this method obtained 98.11% overall accuracy, with a recall of 99.31% for safe code and 93.67% for vulnerable code. Our evaluation utilizes a diverse dataset of over 2,300 CVE-linked and Rust snippets compiled to LLVM IR, facilitating wide-range of coverage across real-world crates.

Medical images and reports offer invaluable insights into patient health. The heterogeneity and complexity of these data hinder effective analysis. To bridge this gap, we investigate contrastive learning models for cross-domain retrieval, which associates medical images with their corresponding clinical reports. This study benchmarks the robustness of four state-of-the-art contrastive learning models: CLIP, CXR-RePaiR, MedCLIP, and CXR-CLIP. We introduce an occlusion retrieval task to evaluate model performance under varying levels of image corruption. Our findings reveal that all evaluated models are highly sensitive to out-of-distribution data, as evidenced by the proportional decrease in performance with increasing occlusion levels. While MedCLIP exhibits slightly more robustness, its overall performance remains significantly behind CXR-CLIP and CXR-RePaiR. CLIP, trained on a general-purpose dataset, struggles with medical image-report retrieval, highlighting the importance of domain-specific training data. The evaluation of this work suggests that more effort needs to be spent on improving the robustness of these models. By addressing these limitations, we can develop more reliable cross-domain retrieval models for medical applications.

This study compared and analyzed the usability, performance, and cost-efficiency of deploying Machine Learning (ML) models in two ML-AI platforms in Google Cloud: BigQuery ML and Vertex AI. Through the experiments with two separate cases, the analysis was conducted with MIMIC-IV datasets of hospitalized patients to deploy regression models on each platform to predict mortality and progression of diseases. The documentation, learning curve, and resource suitability of the platforms were evaluated to access their usability. The study evaluated the total running times and resource utilizations, including storage and compute, to analyze their performance and cost efficiency. The analysis results showed that BigQuery ML offers good usability with easy-to-follow documentation and a moderate learning curve for cloud users, making it more suitable for SQL-savvy users and large-scale data analytics tasks. It also showed efficient resource management and deployment despite its higher initial processing times during the training. Vertex AI incurred higher costs due to longer training times and specific resource allocations. The findings indicate that BigQuery ML seems to be more efficient, particularly in terms of processing time and cost for the experimented clinical dataset and regression models, emphasizing its suitability for large-scale data processing tasks where efficiency is essential.