Karina Yang

Karina Yang, Alexis Bennett, Dominique Duncan

The novel 2019 Coronavirus disease (COVID-19) global pandemic is a defining health crisis. Recent efforts have been increasingly directed towards achieving quick and accurate detection of COVID-19 across symptomatic patients to mitigate the intensity and spread of the disease. Artificial intelligence (AI) algorithms applied to chest X-ray (CXR) images have emerged as promising diagnostic tools, and previous work has demonstrated impressive classification performances. However, such methods have faced criticisms from physicians due to their black-box reasoning process and unpredictable nature. In contrast to professional radiologist diagnosis, AI systems often lack generalizability, explainability, and robustness in the clinical decision making process. In our work, we address these issues by first proposing an extensive baseline study, training and evaluating 21 convolutional neural network (CNN) models on a diverse set of 33,000+ CXR images to classify between healthy, COVID-19, and non-COVID-19 pneumonia CXRs. Our resulting models achieved a 3-way classification accuracy, recall, and precision of up to 97.03\%, 97.97\%, and 99.95\%, respectively. Next, we investigate the effectiveness of adversarial training on model robustness and explainability via Gradient-weighted Class Activation Mapping (Grad-CAM) heatmaps. We find that adversarially trained models not only significantly outperform their standard counterparts on classifying perturbed images, but also yield saliency maps that 1) better specify clinically relevant features, 2) are robust against extraneous artifacts, and 3) agree considerably more with expert radiologist findings.

Bill Yuchen Lin, Yicheng Fu, Karina Yang et al.

We introduce SwiftSage, a novel agent framework inspired by the dual-process theory of human cognition, designed to excel in action planning for complex interactive reasoning tasks. SwiftSage integrates the strengths of behavior cloning and prompting large language models (LLMs) to enhance task completion performance. The framework comprises two primary modules: the Swift module, representing fast and intuitive thinking, and the Sage module, emulating deliberate thought processes. The Swift module is a small encoder-decoder LM fine-tuned on the oracle agent's action trajectories, while the Sage module employs LLMs such as GPT-4 for subgoal planning and grounding. We develop a heuristic method to harmoniously integrate the two modules, resulting in a more efficient and robust problem-solving process. In 30 tasks from the ScienceWorld benchmark, SwiftSage significantly outperforms other methods such as SayCan, ReAct, and Reflexion, demonstrating its effectiveness in solving complex interactive tasks.