A5073573091- Adversarial Robustness in Machine Learning
- Gaussian Processes and Bayesian Inference
- Air Quality Monitoring and Forecasting
- Machine Learning and Data Classification
- Air Quality and Health Impacts
- Statistical Methods and Inference
- Topic Modeling
- Domain Adaptation and Few-Shot Learning
- Anomaly Detection Techniques and Applications
- Fault Detection and Control Systems
- Natural Language Processing Techniques
- Climate Change and Health Impacts
- Advanced Neural Network Applications
- Text Readability and Simplification
- Explainable Artificial Intelligence (XAI)
- Medical Imaging Techniques and Applications
- Hate Speech and Cyberbullying Detection
- Neural Networks and Applications
- Software Engineering Research
- Seismic Imaging and Inversion Techniques
- Child Nutrition and Water Access
- Advanced Multi-Objective Optimization Algorithms
- Advanced Statistical Methods and Models
- Speech and dialogue systems
- Machine Learning in Healthcare
Harvard University
2015-2024
Google (United States)
2020-2024
UC San Diego Health System
2022-2023
Harvard University Press
2019-2023
University of California, Santa Cruz
2023
California University of Pennsylvania
2020
Boston Children's Hospital
2017
Chinese Research Academy of Environmental Sciences
2014
Overconfidence and underconfidence in machine learning classifiers is measured by calibration: the degree to which probabilities predicted for each class match accuracy of classifier on that prediction. How one measures calibration remains a challenge: expected error, most popular metric, has numerous flaws we outline, there no clear empirical understanding how its choices affect conclusions practice, what recommendations are counteract flaws. In this paper, perform comprehensive study...
Growing evidence suggests that fine particulate matter (PM
Bayesian neural networks (BNN) and deep ensembles are principled approaches to estimate the predictive uncertainty of a learning model. However their practicality in real-time, industrial-scale applications limited due heavy memory inference cost. This motivates us study high-quality estimation that require only single network (DNN). By formalizing quantification as minimax problem, we first identify input distance awareness, i.e., model's ability quantify testing example from training data...
Mahalanobis distance (MD) is a simple and popular post-processing method for detecting out-of-distribution (OOD) inputs in neural networks. We analyze its failure modes near-OOD detection propose fix called relative (RMD) which improves performance more robust to hyperparameter choice. On wide selection of challenging vision, language, biology OOD benchmarks (CIFAR-100 vs CIFAR-10, CLINC intent detection, Genomics OOD), we show that RMD meaningfully upon MD (by up 15% AUROC on genomics OOD).
Recent approaches to efficiently ensemble neural networks have shown that strong robustness and uncertainty performance can be achieved with a negligible gain in parameters over the original network. However, these methods still require multiple forward passes for prediction, leading significant computational cost. In this work, we show surprising result: benefits of using predictions `for free' under single model's pass. particular, that, multi-input multi-output (MIMO) configuration, one...
To examine whether a commercial digital health application could support influenza surveillance in China.
Ensemble learning is a standard approach to building machine systems that capture complex phenomena in real-world data. An important aspect of these the complete and valid quantification model uncertainty. We introduce Bayesian nonparametric ensemble (BNE) augments an existing account for different sources BNE model's prediction distribution functions using machinery. It has theoretical guarantee it robustly estimates uncertainty patterns data distribution, can decompose its overall...
Traditional (unstructured) pruning methods for a Transformer model focus on regularizing the individual weights by penalizing them toward zero. In this work, we explore spectral-normalized identity priors (SNIP), structured approach which penalizes an entire residual module in mapping. Our method identifies and discards unimportant non-linear mappings connections applying thresholding operator function norm, is applicable to any including single attention head, blocks, or feed-forward...
Accurate estimation of predictive uncertainty in modern neural networks is critical to achieve well calibrated predictions and detect out-of-distribution (OOD) inputs. The most promising approaches have been predominantly focused on improving model (e.g. deep ensembles Bayesian networks) post-processing techniques for OOD detection ODIN Mahalanobis distance). However, there has relatively little investigation into how the parametrization probabilities discriminative classifiers affects...
Abstract Uncertainty has been a central concept in psychological theories of anxiety. However, this plagued by divergent connotations and operationalizations. The lack consensus hinders the current search for cognitive biological mechanisms anxiety, jeopardizes theory creation comparison, restrains translation basic research into improved diagnoses interventions. Drawing upon uncertainty decomposition Bayesian Decision Theory, we propose well‐defined conceptual structure clinical sciences,...
Structural variants (SVs) play a causal role in numerous diseases but can be difficult to detect and accurately genotype (determine zygosity) with short-read genome sequencing data (SRS). Improving SV genotyping accuracy SRS data, particularly for the many SVs first detected long-read sequencing, will improve our understanding of genetic variation.
Accurate image reconstruction is at the heart of diagnostics in medical imaging. Supervised deep learning-based approaches have been investigated for solving inverse problems including reconstruction. However, these trained models encounter unseen data distributions that are widely shifted from training during deployment. Therefore, it essential to assess whether a given input falls within distribution. Current uncertainty estimation focus on providing an map radiologists, rather than...
Gene-environment and nutrition-environment studies often involve testing of high-dimensional interactions between two sets variables, each having potentially complex nonlinear main effects on an outcome. Construction a valid powerful hypothesis test for such interaction is challenging, due to the difficulty in constructing efficient unbiased estimator complex, effects. In this work we address problem by proposing Cross-validated Ensemble Kernels (CVEK) that learns space appropriate functions...
Content moderation is often performed by a collaboration between humans and machine learning models. However, it not well understood how to design the collaborative process so as maximize combined moderator-model system performance. This work presents rigorous study of this problem, focusing on an approach that incorporates model uncertainty into process. First, we introduce principled metrics describe performance under capacity constraints human moderator, quantifying efficiently utilizes...
Pre-trained seq2seq models excel at graph semantic parsing with rich annotated data, but generalize worse to out-of-distribution (OOD) and long-tail examples. In comparison, symbolic parsers under-perform on population-level metrics, exhibit unique strength in OOD tail generalization. this work, we study compositionality-aware approach neural-symbolic inference informed by model confidence, performing fine-grained reasoning subgraph level (i.e., nodes edges) precisely targeting components...