Uncertainty mitigation refers to the process of reducing or managing uncertainty in a given system. This is often done by identifying potential sources of uncertainty and evaluating the likelihood and potential impact of each risk. Then, strategies can be developed to minimize or manage the risks. The goal of uncertainty mitigation is to make the system more predictable and controllable, and to minimize the potential negative impacts of uncertainty on the system or process.
There are various benefits for uncertainty-driven mitigation. By identifying and managing sources of uncertainty, it can be possible to make a system or process more predictable and controllable. By preparing for a range of potential outcomes through scenario planning and other methods, a system or process can become more resilient to uncertainty and better able to adapt to changing conditions. By quantifying uncertainty and gathering data, it can be possible to make more informed decisions, avoiding costly mistakes or failures..
Leveraging the insights obtained from uncertainty quantification and uncertainty attribution, our research is to develop an uncertainty-driven mitigation strategy to guide the refinement of the model to improve its performance. We develope new techniques under the scope of Bayesian deep learning, as well as to optimize systems and processes in the presence of uncertainty. We are expected to develope new methods for real-time and online uncertainty mitigation, which can help to quickly adapt to changing conditions and prevent negative impacts. We demonstrate the effectiveness of our proposed methods on various computer vision tasks.
Given the uncertainty map generated by uncertainty attribution, we propose a special attention mechanism in the latent space to improve the prediction robustness and reduce uncertainty. We weigh the task-specific features by their uncertainty attributions to strengthen more informative areas, reducing the impacts of the irrelevant background information. We also can filter out the fallacious input information, leading to improved prediction accuracy and robustness.
This approach is useful in situations where certain parts of the input may be more uncertain or less well understood than others, such as when dealing with noisy or incomplete data, or when working with high-dimensional or complex data. By directing the attention of the model towards the most uncertain parts of the input, this approach can help to improve the overall performance and robustness of the model.
With the identified problematic factors and a pre-trained generative model, this research proposes methods to optimize the latent representation of a generative model to generate a better input with less uncertainty.
The idea behind optimizing the latent space is to find the most appropriate representation of the data that can be used to generate new, synthetic data samples that are similar to the original data. By adjusting the parameters of the latent space, it is possible to generate more diverse, realistic, and meaningful data samples that can be used to reduce uncertainty in the model. One one hand, the generated data samples can be used for revising problematic inputs. On the other hand, those new data samples can fill in missing or uncertain parts of the training data. By providing the model with more information and more diverse examples, it can better understand the underlying patterns in the data, and make more accurate predictions.
Imbalanced image classification refers to a problem where the data used for training a classification model is imbalanced, meaning that the number of samples from certain classes is much larger or smaller than the number of samples from other classes. This can lead to a model that is biased towards the more common classes and that has poor performance on the less common classes.
Uncertainty-guided imbalanced image classification is a technique that addresses the problem of imbalanced data by incorporating uncertainty information into the training process. The goal is to use the uncertainty information to guide the model towards the less common classes and to improve the overall performance of the model on imbalanced data.
Uncertainty-guided model fusion is a technique for combining multiple models in order to improve their overall performance and reduce uncertainty. The idea behind this technique is to use the uncertainty information of each model to determine which model is most suitable for a given input and to combine the predictions of the different models in an appropriate way.
Uncertainty-guided curriculum learning aims at leveraging the estimated model uncertainty to dynamically and automatically construct learning curriculum.
Hongji Guo, Nakul Agarwal, Shao-Yuan Lo, Kwonjoon Lee, and Qiang Ji. Uncertainty-aware Action Decoupling Transformer for Action Anticipation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.
Yufei Zhang, Hanjing Wang, Jeffrey O. Kephart, and Qiang Ji. Body Knowledge and Uncertainty Modeling for Monocular 3D Human Body Reconstruction. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. [PDF]
Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-based Spatial-Temporal Attention for Online Action Detection. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. [PDF]
Hongji Guo, Zhou Ren, Yi Wu, Gang Hua, and Qiang Ji. Uncertainty-Guided Probabilistic Transformer for Complex Action Recognition. European Conference on Computer Vision (ECCV), 2022.