Monotonic linear interpolation (MLI) - on the line connecting a random initialization with the minimizer it converges to, the loss and accuracy are monotonic - is a phenomenon that is commonly observed in the training of neural networks. Such a phenomenon may seem to suggest that optimization of neural networks is easy. In this paper, we show that the MLI property is not necessarily related to the hardness of optimization problems, and empirical observations on MLI for deep neural networks depend heavily on biases. In particular, we show that interpolating both weights and biases linearly leads to very different influences on the final output, and when different classes have different last-layer biases on a deep network, there will be a long plateau in both the loss and accuracy interpolation (which existing theory of MLI cannot explain). We also show how the last-layer biases for different classes can be different even on a perfectly balanced dataset using a simple model. Empirically we demonstrate that similar intuitions hold on practical networks and realistic datasets.
Video Question Answering (VideoQA) is the task of answering the natural language questions about a video. Producing an answer requires understanding the interplay across visual scenes in video and linguistic semantics in question. However, most leading VideoQA models work as black boxes, which make the visual-linguistic alignment behind the answering process obscure. Such black-box nature calls for visual explainability that reveals ``What part of the video should the model look at to answer the question?''. Only a few works present the visual explanations in a post-hoc fashion, which emulates the target model's answering process via an additional method. Nonetheless, the emulation struggles to faithfully exhibit the visual-linguistic alignment during answering. Instead of post-hoc explainability, we focus on intrinsic interpretability to make the answering process transparent. At its core is grounding the question-critical cues as the causal scene to yield answers, while rolling out the question-irrelevant information as the environment scene. Taking a causal look at VideoQA, we devise a self-interpretable framework, Equivariant and Invariant Grounding for Interpretable VideoQA (EIGV). Specifically, the equivariant grounding encourages the answering to be sensitive to the semantic changes in the causal scene and question; in contrast, the invariant grounding enforces the answering to be insensitive to the changes in the environment scene. By imposing them on the answering process, EIGV is able to distinguish the causal scene from the environment information, and explicitly present the visual-linguistic alignment. Extensive experiments on three benchmark datasets justify the superiority of EIGV in terms of accuracy and visual interpretability over the leading baselines.
Standard approaches for video recognition usually operate on the full input videos, which is inefficient due to the widely present spatio-temporal redundancy in videos. Recent progress in masked video modelling, i.e., VideoMAE, has shown the ability of vanilla Vision Transformers (ViT) to complement spatio-temporal contexts given only limited visual contents. Inspired by this, we propose propose Masked Action Recognition (MAR), which reduces the redundant computation by discarding a proportion of patches and operating only on a part of the videos. MAR contains the following two indispensable components: cell running masking and bridging classifier. Specifically, to enable the ViT to perceive the details beyond the visible patches easily, cell running masking is presented to preserve the spatio-temporal correlations in videos, which ensures the patches at the same spatial location can be observed in turn for easy reconstructions. Additionally, we notice that, although the partially observed features can reconstruct semantically explicit invisible patches, they fail to achieve accurate classification. To address this, a bridging classifier is proposed to bridge the semantic gap between the ViT encoded features for reconstruction and the features specialized for classification. Our proposed MAR reduces the computational cost of ViT by 53% and extensive experiments show that MAR consistently outperforms existing ViT models with a notable margin. Especially, we found a ViT-Large trained by MAR outperforms the ViT-Huge trained by a standard training scheme by convincing margins on both Kinetics-400 and Something-Something v2 datasets, while our computation overhead of ViT-Large is only 14.5% of ViT-Huge.
For the past 25 years, we have witnessed an extensive application of Machine Learning to the Compiler space; the selection and the phase-ordering problem. However, limited works have been upstreamed into the state-of-the-art compilers, i.e., LLVM, to seamlessly integrate the former into the optimization pipeline of a compiler to be readily deployed by the user. MLGO was among the first of such projects and it only strives to reduce the code size of a binary with an ML-based Inliner using Reinforcement Learning. This paper presents MLGOPerf; the first end-to-end framework capable of optimizing performance using LLVM's ML-Inliner. It employs a secondary ML model to generate rewards used for training a retargeted Reinforcement learning agent, previously used as the primary model by MLGO. It does so by predicting the post-inlining speedup of a function under analysis and it enables a fast training framework for the primary model which otherwise wouldn't be practical. The experimental results show MLGOPerf is able to gain up to 1.8% and 2.2% with respect to LLVM's optimization at O3 when trained for performance on SPEC CPU2006 and Cbench benchmarks, respectively. Furthermore, the proposed approach provides up to 26% increased opportunities to autotune code regions for our benchmarks which can be translated into an additional 3.7% speedup value.
This technical report presents our first place winning solution for temporal action detection task in CVPR-2022 AcitivityNet Challenge. The task aims to localize temporal boundaries of action instances with specific classes in long untrimmed videos. Recent mainstream attempts are based on dense boundary matchings and enumerate all possible combinations to produce proposals. We argue that the generated proposals contain rich contextual information, which may benefits detection confidence prediction. To this end, our method mainly consists of the following three steps: 1) action classification and feature extraction by Slowfast, CSN, TimeSformer, TSP, I3D-flow, VGGish-audio, TPN and ViViT; 2) proposal generation. Our proposed Context-aware Proposal Network (CPN) builds on top of BMN, GTAD and PRN to aggregate contextual information by randomly masking some proposal features. 3) action detection. The final detection prediction is calculated by assigning the proposals with corresponding video-level classifcation results. Finally, we ensemble the results under different feature combination settings and achieve 45.8% performance on the test set, which improves the champion result in CVPR-2021 ActivityNet Challenge by 1.1% in terms of average mAP.
Leading graph contrastive learning (GCL) methods perform graph augmentations in two fashions: (1) randomly corrupting the anchor graph, which could cause the loss of semantic information, or (2) using domain knowledge to maintain salient features, which undermines the generalization to other domains. Taking an invariance look at GCL, we argue that a high-performing augmentation should preserve the salient semantics of anchor graphs regarding instance-discrimination. To this end, we relate GCL with invariant rationale discovery, and propose a new framework, Rationale-aware Graph Contrastive Learning (RGCL). Specifically, without supervision signals, RGCL uses a rationale generator to reveal salient features about graph instance-discrimination as the rationale, and then creates rationale-aware views for contrastive learning. This rationale-aware pre-training scheme endows the backbone model with the powerful representation ability, further facilitating the fine-tuning on downstream tasks. On MNIST-Superpixel and MUTAG datasets, visual inspections on the discovered rationales showcase that the rationale generator successfully captures the salient features (i.e. distinguishing semantic nodes in graphs). On biochemical molecule and social network benchmark datasets, the state-of-the-art performance of RGCL demonstrates the effectiveness of rationale-aware views for contrastive learning. Our codes are available at https://github.com/lsh0520/RGCL.
Bundle recommendation aims to recommend a bundle of related items to users, which can satisfy the users' various needs with one-stop convenience. Recent methods usually take advantage of both user-bundle and user-item interactions information to obtain informative representations for users and bundles, corresponding to bundle view and item view, respectively. However, they either use a unified view without differentiation or loosely combine the predictions of two separate views, while the crucial cooperative association between the two views' representations is overlooked. In this work, we propose to model the cooperative association between the two different views through cross-view contrastive learning. By encouraging the alignment of the two separately learned views, each view can distill complementary information from the other view, achieving mutual enhancement. Moreover, by enlarging the dispersion of different users/bundles, the self-discrimination of representations is enhanced. Extensive experiments on three public datasets demonstrate that our method outperforms SOTA baselines by a large margin. Meanwhile, our method requires minimal parameters of three set of embeddings (user, bundle, and item) and the computational costs are largely reduced due to more concise graph structure and graph learning module. In addition, various ablation and model studies demystify the working mechanism and justify our hypothesis. Codes and datasets are available at https://github.com/mysbupt/CrossCBR.
Video Question Answering (VideoQA) is the task of answering questions about a video. At its core is understanding the alignments between visual scenes in video and linguistic semantics in question to yield the answer. In leading VideoQA models, the typical learning objective, empirical risk minimization (ERM), latches on superficial correlations between video-question pairs and answers as the alignments. However, ERM can be problematic, because it tends to over-exploit the spurious correlations between question-irrelevant scenes and answers, instead of inspecting the causal effect of question-critical scenes. As a result, the VideoQA models suffer from unreliable reasoning. In this work, we first take a causal look at VideoQA and argue that invariant grounding is the key to ruling out the spurious correlations. Towards this end, we propose a new learning framework, Invariant Grounding for VideoQA (IGV), to ground the question-critical scene, whose causal relations with answers are invariant across different interventions on the complement. With IGV, the VideoQA models are forced to shield the answering process from the negative influence of spurious correlations, which significantly improves the reasoning ability. Experiments on three benchmark datasets validate the superiority of IGV in terms of accuracy, visual explainability, and generalization ability over the leading baselines.