Fine-tuning foundation models often compromises their robustness to distribution shifts. To remedy this, most robust fine-tuning methods aim to preserve the pre-trained features. However, not all pre-trained features are robust and those methods are largely indifferent to which ones to preserve. We propose dual risk minimization (DRM), which combines empirical risk minimization with worst-case risk minimization, to better preserve the core features of downstream tasks. In particular, we utilize core-feature descriptions generated by LLMs to induce core-based zero-shot predictions which then serve as proxies to estimate the worst-case risk. DRM balances two crucial aspects of model robustness: expected performance and worst-case performance, establishing a new state of the art on various real-world benchmarks. DRM significantly improves the out-of-distribution performance of CLIP ViT-L/14@336 on ImageNet (75.9$\to$77.1), WILDS-iWildCam (47.1$\to$51.8), and WILDS-FMoW (50.7$\to$53.1); opening up new avenues for robust fine-tuning. Our code is available at https://github.com/vaynexie/DRM.

Micro-expressions (MEs) are involuntary and subtle facial expressions that are thought to reveal feelings people are trying to hide. ME spotting detects the temporal intervals containing MEs in videos. Detecting such quick and subtle motions from long videos is difficult. Recent works leverage detailed facial motion representations, such as the optical flow, and deep learning models, leading to high computational complexity. To reduce computational complexity and achieve real-time operation, we propose RMES, a real-time ME spotting framework. We represent motion using phase computed by Riesz Pyramid, and feed this motion representation into a three-stream shallow CNN, which predicts the likelihood of each frame belonging to an ME. In comparison to optical flow, phase provides more localized motion estimates, which are essential for ME spotting, resulting in higher performance. Using phase also reduces the required computation of the ME spotting pipeline by 77.8%. Despite its relative simplicity and low computational complexity, our framework achieves state-of-the-art performance on two public datasets: CAS(ME)2and SAMM Long Videos.

Model explanations are generated by XAI (explainable AI) methods to help people understand and interpret machine learning models. To study XAI methods from the human perspective, we propose a human-based benchmark dataset, i.e., human saliency benchmark (HSB), for evaluating saliency-based XAI methods. Different from existing human saliency annotations where class-related features are manually and subjectively labeled, this benchmark collects more objective human attention on vision information with a precise eye-tracking device and a novel crowdsourcing experiment. Taking the labor cost of human experiment into consideration, we further explore the potential of utilizing a prediction model trained on HSB to mimic saliency annotating by humans. Hence, a dense prediction problem is formulated, and we propose an encoder-decoder architecture which combines multi-modal and multi-scale features to produce the human saliency maps. Accordingly, a pretraining-finetuning method is designed to address the model training problem. Finally, we arrive at a model trained on HSB named human saliency imitator (HSI). We show, through an extensive evaluation, that HSI can successfully predict human saliency on our HSB dataset, and the HSI-generated human saliency dataset on ImageNet showcases the ability of benchmarking XAI methods both qualitatively and quantitatively.

Given a fixed parameter budget, one can build a single large neural network or create a memory-split ensemble: a pool of several smaller networks with the same total parameter count as the single network. A memory-split ensemble can outperform its single model counterpart (Lobacheva et al., 2020): a phenomenon known as the memory-split advantage (MSA). The reasons for MSA are still not yet fully understood. In particular, it is difficult in practice to predict when it will exist. This paper sheds light on the reasons underlying MSA using random feature theory. We study the dependence of the MSA on several factors: the parameter budget, the training set size, the L2 regularization and the Stochastic Gradient Descent (SGD) hyper-parameters. Using the bias-variance decomposition, we show that MSA exists when the reduction in variance due to the ensemble (\ie, \textit{ensemble gain}) exceeds the increase in squared bias due to the smaller size of the individual networks (\ie, \textit{shrinkage cost}). Taken together, our theoretical analysis demonstrates that the MSA mainly exists for the small parameter budgets relative to the training set size, and that memory-splitting can be understood as a type of regularization. Adding other forms of regularization, \eg L2 regularization, reduces the MSA. Thus, the potential benefit of memory-splitting lies primarily in the possibility of speed-up via parallel computation. Our empirical experiments with deep neural networks and large image datasets show that MSA is not a general phenomenon, but mainly exists when the number of training iterations is small.

Spatial-temporal feature learning is of vital importance for video emotion recognition. Previous deep network structures often focused on macro-motion which extends over long time scales, e.g., on the order of seconds. We believe integrating structures capturing information about both micro- and macro-motion will benefit emotion prediction, because human perceive both micro- and macro-expressions. In this paper, we propose to combine micro- and macro-motion features to improve video emotion recognition with a two-stream recurrent network, named MIMAMO (Micro-Macro-Motion) Net. Specifically, smaller and shorter micro-motions are analyzed by a two-stream network, while larger and more sustained macro-motions can be well captured by a subsequent recurrent network. Assigning specific interpretations to the roles of different parts of the network enables us to make choice of parameters based on prior knowledge: choices that turn out to be optimal. One of the important innovations in our model is the use of interframe phase differences rather than optical flow as input to the temporal stream. Compared with the optical flow, phase differences require less computation and are more robust to illumination changes. Our proposed network achieves state of the art performance on two video emotion datasets, the OMG emotion dataset and the Aff-Wild dataset. The most significant gains are for arousal prediction, for which motion information is intuitively more informative. Source code is available at https://github.com/wtomin/MIMAMO-Net.

We train a unified model to perform three tasks: facial action unit detection, expression classification, and valence-arousal estimation. We address two main challenges of learning the three tasks. First, most existing datasets are highly imbalanced. Second, most existing datasets do not contain labels for all three tasks. To tackle the first challenge, we apply data balancing techniques to experimental datasets. To tackle the second challenge, we propose an algorithm for the multitask model to learn from missing (incomplete) labels. This algorithm has two steps. We first train a teacher model to perform all three tasks, where each instance is trained by the ground truth label of its corresponding task. Secondly, we refer to the outputs of the teacher model as the soft labels. We use the soft labels and the ground truth to train the student model. We find that most of the student models outperform their teacher model on all the three tasks. Finally, we use model ensembling to boost performance further on the three tasks. Our code is publicly available1.1https://github.com/wtomin/multitask-Emotion-Recognition-withIncomplete-Labels