Video Understanding

We present Generalizable NeRF Transformer (GNT), a transformer-based architecture that reconstructs Neural Radiance Fields (NeRFs) and learns to render novel views on the fly from source views. While prior works on NeRFs optimize a scene representation by inverting a handcrafted rendering equation, GNT achieves neural representation and rendering that generalizes across scenes using transformers at two stages. (1) The view transformer leverages multi-view geometry as an inductive bias for attention-based scene representation, and predicts coordinate-aligned features by aggregating information from epipolar lines on the neighboring views. (2) The ray transformer renders novel views using attention to decode the features from the view transformer along the sampled points during ray marching. Our experiments demonstrate that when optimized on a single scene, GNT can successfully reconstruct NeRF without an explicit rendering formula due to the learned ray renderer. When trained on multiple scenes, GNT consistently achieves state-of-the-art performance when transferring to unseen scenes and outperform all other methods by ~10% on average. Our analysis of the learned attention maps to infer depth and occlusion indicate that attention enables learning a physically-grounded rendering. Our results show the promise of transformers as a universal modeling tool for graphics. Please refer to our project page for video results: https://vita-group.github.io/GNT/

Crossmodal knowledge distillation (KD) extends traditional knowledge distillation to the area of multimodal learning and demonstrates great success in various applications. To achieve knowledge transfer across modalities, a pretrained network from one modality is adopted as the teacher to provide supervision signals to a student network learning from the other modality. In contrast to the empirical success reported in prior works, the working mechanism of crossmodal KD remains a mystery. In this paper, we present a thorough understanding of crossmodal KD. We begin by providing two failure cases and demonstrate that KD is not a universal cure in crossmodal knowledge transfer. We then present the modality Venn diagram to understand modality relationships and the modality focusing hypothesis revealing the decisive factor in the efficacy of crossmodal KD. Experimental results on 6 multimodal datasets help justify our hypothesis, diagnose failure cases, and point directions to improve crossmodal knowledge transfer in the future.

Unsupervised domain adaptation for video recognition is challenging where the domain shift includes both spatial variations and temporal dynamics. Previous works have focused on exploring contrastive learning for cross-domain alignment. However, limited variations in intra-domain positives, false cross-domain positives, and false negatives hinder contrastive learning from fulfilling intra-domain discrimination and cross-domain closeness. This paper presents a non-contrastive learning framework without relying on negative samples for unsupervised video domain adaptation. To address the limited variations in intra-domain positives, we set unlabeled target videos as anchors and explored to mine "informative intra-domain positives" in the form of spatial/temporal augmentations and target nearest neighbors (NNs). To tackle the false cross-domain positives led by noisy pseudo-labels, we reversely set source videos as anchors and sample the synthesized target videos as "robust cross-domain positives" from an estimated target distribution, which are naturally more robust to the pseudo-label noise. Our approach is demonstrated to be superior to state-of-the-art methods through extensive experiments on several cross-domain action recognition benchmarks.

We study embodied reference understanding, the task of locating referents using embodied gestural signals and language references. Human studies have revealed that, contrary to popular belief, objects referred to or pointed to do not lie on the elbow-wrist line, but rather on the so-called virtual touch line. Nevertheless, contemporary human pose representations lack the virtual touch line. To tackle this problem, we devise the touch-line Transformer: It takes as input tokenized visual and textual features and simultaneously predicts the referent’s bounding box and a touch-line vector. Leveraging this touch-line prior, we further devise a geometric consistency loss that promotes co-linearity between referents and touch lines. Using the touch line as gestural information dramatically improves model performances: Experiments on the YouRefIt dataset demonstrate that our method yields a +25.0% accuracy improvement under the 0.75 IoU criterion, hence closing 63.6% of the performance difference between models and humans. Furthermore, we computationally validate prior human studies by demonstrating that computational models more accurately locate referents when employing the virtual touch line than when using the elbow-wrist line.

Binary Neural Networks (BNNs) are compact and efficient by using binary weights instead of real-valued weights. Current BNNs use latent real-valued weights during training, where several training hyper-parameters are inherited from real-valued networks. The interpretation of several of these hyperparameters is based on the magnitude of the real-valued weights. For BNNs, however, the magnitude of binary weights is not meaningful, and thus it is unclear what these hyperparameters actually do. One example is weight-decay, which aims to keep the magnitude of real-valued weights small. Other examples are latent weight initialization, the learning rate, and learning rate decay, which influence the magnitude of the real-valued weights. The magnitude is interpretable for real-valued weights, but loses its meaning for binary weights. In this paper we offer a new interpretation of these magnitude-based hyperparameters based on higher-order gradient filtering during network optimization. Our analysis makes it possible to understand how magnitude-based hyperparameters influence the training of binary networks which allows for new optimization filters specifically designed for binary neural networks that are independent of their real-valued interpretation. Moreover, our improved understanding reduces the number of hyperparameters, which in turn eases the hyperparameter tuning effort which may lead to better hyperparameter values for improved accuracy. Code is available at https://github.com/jorisquist/Understanding-WM-HP-in-BNNs

When trained on language data, do transformers learn some arbitrary computation that utilizes the full capacity of the architecture or do they learn a simpler, tree-like computation, hypothesized to underlie compositional meaning systems like human languages? There is an apparent tension between compositional accounts of human language understanding, which are based on a restricted bottom-up computational process, and the enormous success of neural models like transformers, which can route information arbitrarily between different parts of their input. One possibility is that these models, while extremely flexible in principle, in practice learn to interpret language hierarchically, ultimately building sentence representations close to those predictable by a bottom-up, tree-structured model. To evaluate this possibility, we describe an unsupervised and parameter-free method to \emph{functionally project} the behavior of any transformer into the space of tree-structured networks. Given an input sentence, we produce a binary tree that approximates the transformer's representation-building process and a score that captures how ``tree-like'' the transformer's behavior is on the input. While calculation of this score does not require training any additional models, it provably upper-bounds the fit between a transformer and any tree-structured approximation. Using this method, we show that transformers for three different tasks become more tree-like over the course of training, in some cases unsupervisedly recovering the same trees as supervised parsers. These trees, in turn, are predictive of model behavior, with more tree-like models generalizing better on tests of compositional generalization.

While the empirical success of self-supervised learning (SSL) heavily relies on the usage of deep nonlinear models, existing theoretical works on SSL understanding still focus on linear ones. In this paper, we study the role of nonlinearity in the training dynamics of contrastive learning (CL) on one and two-layer nonlinear networks with homogeneous activation $h(x) = h'(x)x$. We have two major theoretical discoveries. First, the presence of nonlinearity can lead to many local optima even in 1-layer setting, each corresponding to certain patterns from the data distribution, while with linear activation, only one major pattern can be learned. This suggests that models with lots of parameters can be regarded as a \emph{brute-force} way to find these local optima induced by nonlinearity. Second, in the 2-layer case, linear activation is proven not capable of learning specialized weights into diverse patterns, demonstrating the importance of nonlinearity. In addition, for 2-layer setting, we also discover \emph{global modulation}: those local patterns discriminative from the perspective of global-level patterns are prioritized to learn, further characterizing the learning process. Simulation verifies our theoretical findings.

Diffusion models have been recently employed to improve certified robustness through the process of denoising. However, the theoretical understanding of why diffusion models are able to improve the certified robustness is still lacking, preventing from further improvement. In this study, we close this gap by analyzing the fundamental properties of diffusion models and establishing the conditions under which they can enhance certified robustness. This deeper understanding allows us to propose a new method DensePure, designed to improve the certified robustness of a pretrained model (i.e. classifier). Given an (adversarial) input, DensePure consists of multiple runs of denoising via the reverse process of the diffusion model (with different random seeds) to get multiple reversed samples, which are then passed through the classifier, followed by majority voting of inferred labels to make the final prediction. This design of using multiple runs of denoising is informed by our theoretical analysis of the conditional distribution of the reversed sample. Specifically, when the data density of a clean sample is high, its conditional density under the reverse process in a diffusion model is also high; thus sampling from the latter conditional distribution can purify the adversarial example and return the corresponding clean sample with a high probability. By using the highest density point in the conditional distribution as the reversed sample, we identify the robust region of a given instance under the diffusion model's reverse process. We show that this robust region is a union of multiple convex sets, and is potentially much larger than the robust regions identified in previous works. In practice, DensePure can approximate the label of the high density region in the conditional distribution so that it can enhance certified robustness. We conduct extensive experiments to demonstrate the effectiveness of DensePure by evaluating its certified robustness given a standard model via randomized smoothing. We show that DensePure is consistently better than existing methods on ImageNet, with 7% improvement on average.

Recently, researchers observed that gradient descent for deep neural networks operates in an ``edge-of-stability'' (EoS) regime: the sharpness (maximum eigenvalue of the Hessian) is often larger than stability threshold $2/\eta$ (where $\eta$ is the step size). Despite this, the loss oscillates and converges in the long run, and the sharpness at the end is just slightly below $2/\eta$. While many other well-understood nonconvex objectives such as matrix factorization or two-layer networks can also converge despite large sharpness, there is often a larger gap between sharpness of the endpoint and $2/\eta$. In this paper, we study EoS phenomenon by constructing a simple function that has the same behavior. We give rigorous analysis for its training dynamics in a large local region and explain why the final converging point has sharpness close to $2/\eta$. Globally we observe that the training dynamics for our example has an interesting bifurcating behavior, which was also observed in the training of neural nets.

3D convolution neural networks (CNNs) have been the prevailing option for video recognition. To capture the temporal information, 3D convolutions are computed along the sequences, leading to cubically growing and expensive computations. To reduce the computational cost, previous methods resort to manually designed 3D/2D CNN structures with approximations or automatic search, which sacrifice the modeling ability or make training time-consuming. In this work, we propose to automatically design efficient 3D CNN architectures via a novel training-free neural architecture search approach tailored for 3D CNNs considering the model complexity. To measure the expressiveness of 3D CNNs efficiently, we formulate a 3D CNN as an information system and derive an analytic entropy score, based on the Maximum Entropy Principle. Specifically, we propose a spatio-temporal entropy score (STEntr-Score) with a refinement factor to handle the discrepancy of visual information in spatial and temporal dimensions, through dynamically leveraging the correlation between the feature map size and kernel size depth-wisely. Highly efficient and expressive 3D CNN architectures, i.e., entropy-based 3D CNNs (E3D family), can then be efficiently searched by maximizing the STEntr-Score under a given computational budget, via an evolutionary algorithm without training the network parameters. Extensive experiments on Something-Something V1&V2 and Kinetics400 demonstrate that the E3D family achieves state-of-the-art performance with higher computational efficiency.

Vision Transformers (ViTs) with self-attention modules have recently achieved great empirical success in many vision tasks. Due to non-convex interactions across layers, however, the theoretical learning and generalization analysis is mostly elusive. Based on a data model characterizing both label-relevant and label-irrelevant tokens, this paper provides the first theoretical analysis of training a three-layer ViT, i.e., one self-attention layer followed by a two-layer perceptron, for a classification task. We characterize the sample complexity to achieve a zero generalization error. Our sample complexity bound is positively correlated with the inverse of the fraction of label-relevant tokens, the token noise level, and the initial model error. We also prove that a training process using stochastic gradient descent (SGD) leads to a sparse attention map, which is a formal verification of the general intuition about the success of attention. Moreover, this paper indicates that a proper token sparsification can improve the test performance by removing label-irrelevant and/or noisy tokens, including spurious correlations. Empirical experiments on synthetic data and CIFAR-10 dataset justify our theoretical results and generalize to deeper ViTs.

Deep learning video compression outperforms its hand-craft counterparts with enhanced flexibility and capacity. One key component of the learned video codec is the autoregressive entropy model conditioned on spatial and temporal priors. Operating autoregressive on raster scanning order naively treats the context as unidirectional. This is neither efficient nor optimal, considering that conditional information probably locates at the end of the sequence. We thus introduce an entropy model based on a masked image modeling transformer (MIMT) to learn the spatial-temporal dependencies. Video frames are first encoded into sequences of tokens and then processed with the transformer encoder as priors. The transformer decoder learns the probability mass functions (PMFs) \emph{conditioned} on the priors and masked inputs. Then it is capable of selecting optimal decoding orders without a fixed direction. During training, MIMT aims to predict the PMFs of randomly masked tokens by attending to tokens in all directions. This allows MIMT to capture the temporal dependencies from encoded priors and the spatial dependencies from the unmasked tokens, i.e., decoded tokens. At inference time, the model begins with generating PMFs of all masked tokens in parallel and then decodes the frame iteratively from the previously-selected decoded tokens (i.e., with high confidence). In addition, we improve the overall performance with more techniques, e.g., manifold conditional priors accumulating a long range of information, shifted window attention to reduce complexity. Extensive experiments demonstrate the proposed MIMT framework equipped with the new transformer entropy model achieves state-of-the-art performance on HEVC, UVG, and MCL-JCV datasets, generally outperforming the VVC in terms of PSNR and SSIM.

Adaptive gradient methods such as Adam have gained increasing popularity in deep learning optimization. However, it has been observed in many deep learning applications such as image classification, Adam can converge to a different solution with a worse test error compared to (stochastic) gradient descent, even with a fine-tuned regularization. In this paper, we provide a theoretical explanation for this phenomenon: we show that in the nonconvex setting of learning over-parameterized two-layer convolutional neural networks starting from the same random initialization, for a class of data distributions (inspired from image data), Adam and gradient descent (GD) can converge to different global solutions of the training objective with provably different generalization errors, even with weight decay regularization. In contrast, we show that if the training objective is convex, and the weight decay regularization is employed, any optimization algorithms including Adam and GD will converge to the same solution if the training is successful. This suggests that the generalization gap between Adam and SGD in the presence of weight decay regularization is closely tied to the nonconvex landscape of deep learning optimization, which cannot be covered by the recent neural tangent kernel (NTK) based analysis.

The promise of multimodal models for real-world applications has inspired research in visualizing and understanding their internal mechanics with the end goal of empowering stakeholders to visualize model behavior, perform model debugging, and promote trust in machine learning models. However, modern multimodal models are typically black-box neural networks, which makes it challenging to understand their internal mechanics. How can we visualize the internal modeling of multimodal interactions in these models? Our paper aims to fill this gap by proposing MultiViz, a method for analyzing the behavior of multimodal models by scaffolding the problem of interpretability into 4 stages: (1) unimodal importance: how each modality contributes towards downstream modeling and prediction, (2) cross-modal interactions: how different modalities relate with each other, (3) multimodal representations: how unimodal and cross-modal interactions are represented in decision-level features, and (4) multimodal prediction: how decision-level features are composed to make a prediction. MultiViz is designed to operate on diverse modalities, models, tasks, and research areas. Through experiments on 8 trained models across 6 real-world tasks, we show that the complementary stages in MultiViz together enable users to (1) simulate model predictions, (2) assign interpretable concepts to features, (3) perform error analysis on model misclassifications, and (4) use insights from error analysis to debug models. MultiViz is publicly available, will be regularly updated with new interpretation tools and metrics, and welcomes inputs from the community.

Data augmentation has recently emerged as an essential component of modern training recipes for visual recognition tasks. However, data augmentation for video recognition has been rarely explored despite its effectiveness. Few existing augmentation recipes for video recognition naively extend the image augmentation methods by applying the same operations to the whole video frames. Our main idea is that the magnitude of augmentation operations for each frame needs to be changed over time to capture the real-world video's temporal variations. These variations should be generated as diverse as possible using fewer additional hyper-parameters during training. Through this motivation, we propose a simple yet effective video data augmentation framework, DynaAugment. The magnitude of augmentation operations on each frame is changed by an effective mechanism, Fourier Sampling that parameterizes diverse, smooth, and realistic temporal variations. DynaAugment also includes an extended search space suitable for video for automatic data augmentation methods. DynaAugment experimentally demonstrates that there are additional performance rooms to be improved from static augmentations on diverse video models. Specifically, we show the effectiveness of DynaAugment on various video datasets and tasks: large-scale video recognition (Kinetics-400 and Something-Something-v2), small-scale video recognition (UCF-101 and HMDB-51), fine-grained video recognition (Diving-48 and FineGym), video action segmentation on Breakfast, video action localization on THUMOS'14, and video object detection on MOT17Det.

Influence functions estimate effect of individual data points on predictions of the model on test data and were adapted to deep learning in \cite{koh2017understanding}. They have been used for detecting data poisoning, detecting helpful and harmful examples, influence of groups of datapoints, etc. Recently, \cite{ilyas2022datamodels} introduced a linear regression method they termed {\em datamodels} to predict the effect of training points on outputs on test data. The current paper seeks to provide a better theoretical understanding of such interesting empirical phenomena. The primary tool is harmonic analysis and the idea of {\em noise stability}. Contributions include: (a) Exact characterization of the learnt datamodel in terms of Fourier coefficients. (b) An efficient method to estimate the residual error and quality of the optimum linear datamodel without having to train the datamodel. (c) New insights into when influences of groups of datapoints may or may not add up linearly.

Recently, a variety of methods under the name of non-contrastive learning (like BYOL, SimSiam, SwAV, DINO) show that when equipped with some asymmetric architectural designs, aligning positive pairs alone is sufficient to attain good performance in self-supervised visual learning. Despite some understandings of some specific modules (like the predictor in BYOL), there is yet no unified theoretical understanding of how these seemingly different asymmetric designs can all avoid feature collapse, particularly considering methods that also work without the predictor (like DINO). In this work, we propose a unified theoretical understanding for existing variants of non-contrastive learning. Our theory named Rank Differential Mechanism (RDM) shows that all these asymmetric designs create a consistent rank difference in their dual-branch output features. This rank difference will provably lead to an improvement of effective dimensionality and alleviate either complete or dimensional feature collapse. Different from previous theories, our RDM theory is applicable to different asymmetric designs (with and without the predictor), and thus can serve as a unified understanding of existing non-contrastive learning methods. Besides, our RDM theory also provides practical guidelines for designing many new non-contrastive variants. We show that these variants indeed achieve comparable performance to existing methods on benchmark datasets, and some of them even outperform the baselines. Our code is available at \url{https://github.com/PKU-ML/Rank-Differential-Mechanism}.

Empirical risk minimization (ERM) is known to be non-robust in practice to distributional shift where the training and the test distributions are different. A suite of approaches, such as importance weighting, and variants of distributionally robust optimization (DRO), have been proposed to solve this problem. But a line of recent work has empirically shown that these approaches do not significantly improve over ERM in real applications with distribution shift. The goal of this work is to obtain a comprehensive theoretical understanding of this intriguing phenomenon. We first posit the class of Generalized Reweighting (GRW) algorithms, as a broad category of approaches that iteratively update model parameters based on iterative reweighting of the training samples. We show that when overparameterized models are trained under GRW, the resulting models are close to that obtained by ERM. We also show that adding small regularization which does not greatly affect the empirical training accuracy does not help. Together, our results show that a broad category of what we term GRW approaches are not able to achieve distributionally robust generalization. Our work thus has the following sobering takeaway: to make progress towards distributionally robust generalization, we either have to develop non-GRW approaches, or perhaps devise novel classification/regression loss functions that are adapted to GRW approaches.

Contrastive learning, especially self-supervised contrastive learning (SSCL), has achieved great success in extracting powerful features from unlabeled data. In this work, we contribute to the theoretical understanding of SSCL and uncover its connection to the classic data visualization method, stochastic neighbor embedding (SNE), whose goal is to preserve pairwise distances. From the perspective of preserving neighboring information, SSCL can be viewed as a special case of SNE with the input space pairwise similarities specified by data augmentation. The established correspondence facilitates deeper theoretical understanding of learned features of SSCL, as well as methodological guidelines for practical improvement. Specifically, through the lens of SNE, we provide novel analysis on domain-agnostic augmentations, implicit bias and robustness of learned features. To illustrate the practical advantage, we demonstrate that the modifications from SNE to $t$-SNE can also be adopted in the SSCL setting, achieving significant improvement in both in-distribution and out-of-distribution generalization.

We are interested in solving a class of problems that seek to understand and adopt rational behavior from demonstrations. We may broadly classify these problems into four categories of reward identification, counterfactual analysis, behavior imitation, and behavior transfer. In this work, we make a key observation that knowing how changes in the underlying rewards affect the optimal behavior allows one to solve a variety of aforementioned problems. To a local approximation, this quantity is precisely captured by what we term the Bellman score, i.e gradient of log probabilities of the optimal policy with respect to the reward. We introduce the Bellman score operator which provably converges to the gradient of the infinite-horizon optimal Q-values with respect to the reward which can then be used to directly estimate the score. Guided by our theory, we derive a practical score-learning algorithm which can be used for score estimation in high-dimensional state-actions spaces. We show that score-learning can be used to reliably identify rewards, perform counterfactual predictions, achieve state-of-the-art behavior imitation, and transfer policies across environments.