Minisymposium

Advancements in remote sensing, geospatial data, physicochemical and source apportionment models, citizen science networks and machine learning have greatly improved our ability to predict air pollution at high spatiotemporal resolution and over large domains and time periods. Air pollution models with high predictive performance and low uncertainty are critical for estimating population and individual level human exposures and their health risks, in retrospective studies and for forecasting. This talk will present advances in air pollution modeling that integrate multi-modal data and deep learning and estimate uncertainty in predictions which can then inform or be integrated into epidemiological analyses. Current challenges and future needs for advancing these models that are currently being pursued by our team for biopreparedness and health risk applications will also be discussed.

A deep abstaining classifier, or DAC, introduced first for combating label noise, is a regular deep neural network classifier (DNN) but with an additional (abstain) class and a custom loss function that permits abstention during training. This allows the DAC to identify and abstain on (or decline to classify) confusing samples, without the need for manually labeling these cases, while continuing to learn and improve classification performance on the non-abstained samples. It has been shown that the resulting models can significantly improve accuracy compared to the original DNN, at the cost of reduced coverage. The DAC learns patterns in the data that make prediction unreliable, is more robust to feature noise, and constitutes a useful tool to diagnose uncertain predictions at inference time. In this talk, we describe how we adapt the DAC to be combined with Federated Learning (FL) to allow for a distributed training using data from different silos that cannot be openly shared due to privacy concerns, as is frequently encountered when dealing with health records. We also demonstrate how this DAC+FL model can be applied for anomaly detection and discuss how to configure this per silo.

High-quality, accurate, and real-time information about disease spread is critical for rapid response to a biothreat. Electronic health data is collected by approximately 90% of all physicians in the United States. However, its broad use for public health surveillance and monitoring is inhibited by a lack of organization around common and high-quality information. We use AI to automatically code unstructured clinical documents. During a biothreat scenario, these tools will help synthesize data across multiple medical records systems and enable the rapid identification of vulnerable populations. In addition to extracting important information from clinical records, we are developing an autonomous biothreat agent that scans reputable public health surveillance reports and identifies emerging threats. Once an emerging threat is identified, our AI agents will be able to search existing scientific literature to extract the disease specific parameters for epidemiological modeling. AI is helping us develop the necessary infrastructure for near real-time situational readiness, so that we will be prepared during the next Covid-like event.

As Large Language Models (LLMs) grow in size and capability, their high computational cost poses a major challenge for real-time applications, making efficient inference a critical research problem. Speculative Decoding (SD) has emerged as a promising technique to accelerate LLM inference by leveraging a smaller draft model to generate candidate tokens, which are then verified in parallel by a larger target model to ensure statistical consistency. However, the need for frequent verification calls to the target LLM limits the potential speedup of SD. We propose SPRINTER, which utilizes a low-complexity verifier trained to predict if tokens generated by the draft model would be accepted by the target LLM. By performing approximate sequential verification, SPRINTER eliminates the need for constant verification by the target LLM and is only invoked when a token is deemed unacceptable. This significantly reduces the number of calls to the larger model, enabling further acceleration. We present a theoretical analysis of SPRINTER, examining the statistical properties of the generated tokens and the expected reduction in latency as a function of the verifier. Our evaluations on multiple datasets and model pairs demonstrate that approximate verification can maintain high-quality generation while achieving even greater speedups.