Junjie Yang

Junjie Yang, Tianlong Chen, Xuxi Chen et al.

Neural networks have demonstrated success in various domains, yet their performance can be significantly degraded by even a small input perturbation. Consequently, the construction of such perturbations, known as adversarial attacks, has gained significant attention, many of which fall within "white-box" scenarios where we have full access to the neural network. Existing attack algorithms, such as the projected gradient descent (PGD), commonly take the sign function on the raw gradient before updating adversarial inputs, thereby neglecting gradient magnitude information. In this paper, we present a theoretical analysis of how such sign-based update algorithm influences step-wise attack performance, as well as its caveat. We also interpret why previous attempts of directly using raw gradients failed. Based on that, we further propose a new raw gradient descent (RGD) algorithm that eliminates the use of sign. Specifically, we convert the constrained optimization problem into an unconstrained one, by introducing a new hidden variable of non-clipped perturbation that can move beyond the constraint. The effectiveness of the proposed RGD algorithm has been demonstrated extensively in experiments, outperforming PGD and other competitors in various settings, without incurring any additional computational overhead. The codes is available in https://github.com/JunjieYang97/RGD.

Pengfei Wang, Guohai Xu, Weinong Wang et al.

Recent advancements have enhanced the capability of Multimodal Large Language Models (MLLMs) to comprehend multi-image information. However, existing benchmarks primarily evaluate answer correctness, overlooking whether models genuinely comprehend the visual input. To address this, we define implicit visual misunderstanding (IVM), where MLLMs provide correct answers without fully comprehending the visual input. Through our analysis, we decouple the visual and textual modalities within the causal attention module, revealing that attention distribution increasingly converges on the image associated with the correct answer as the network layers deepen. This insight leads to the introduction of a scale-agnostic metric, \textit{attention accuracy}, and a novel benchmark for quantifying IVMs. Attention accuracy directly evaluates the model's visual understanding via internal mechanisms, remaining robust to positional biases for more reliable assessments. Furthermore, we extend our approach to finer granularities and demonstrate its effectiveness in unimodal scenarios, underscoring its versatility and generalizability.

Tamim El Ahmad, Junjie Yang, Pierre Laforgue et al.

By leveraging the kernel trick in the output space, kernel-induced losses provide a principled way to define structured output prediction tasks for a wide variety of output modalities. In particular, they have been successfully used in the context of surrogate non-parametric regression, where the kernel trick is typically exploited in the input space as well. However, when inputs are images or texts, more expressive models such as deep neural networks seem more suited than non-parametric methods. In this work, we tackle the question of how to train neural networks to solve structured output prediction tasks, while still benefiting from the versatility and relevance of kernel-induced losses. We design a novel family of deep neural architectures, whose last layer predicts in a data-dependent finite-dimensional subspace of the infinite-dimensional output feature space deriving from the kernel-induced loss. This subspace is chosen as the span of the eigenfunctions of a randomly-approximated version of the empirical kernel covariance operator. Interestingly, this approach unlocks the use of gradient descent algorithms (and consequently of any neural architecture) for structured prediction. Experiments on synthetic tasks as well as real-world supervised graph prediction problems show the relevance of our method.

Junjie Yang, Kaiyi Ji, Yingbin Liang

Bilevel optimization has been widely applied in many important machine learning applications such as hyperparameter optimization and meta-learning. Recently, several momentum-based algorithms have been proposed to solve bilevel optimization problems faster. However, those momentum-based algorithms do not achieve provably better computational complexity than $\mathcal{\widetilde O}(ε^{-2})$ of the SGD-based algorithm. In this paper, we propose two new algorithms for bilevel optimization, where the first algorithm adopts momentum-based recursive iterations, and the second algorithm adopts recursive gradient estimations in nested loops to decrease the variance. We show that both algorithms achieve the complexity of $\mathcal{\widetilde O}(ε^{-1.5})$, which outperforms all existing algorithms by the order of magnitude. Our experiments validate our theoretical results and demonstrate the superior empirical performance of our algorithms in hyperparameter applications.

Cheng Chen, Junjie Yang, Yi Zhou

Modern deep neural network (DNN) trainings utilize various training techniques, e.g., nonlinear activation functions, batch normalization, skip-connections, etc. Despite their effectiveness, it is still mysterious how they help accelerate DNN trainings in practice. In this paper, we provide an empirical study of the regularization effect of these training techniques on DNN optimization. Specifically, we find that the optimization trajectories of successful DNN trainings consistently obey a certain regularity principle that regularizes the model update direction to be aligned with the trajectory direction. Theoretically, we show that such a regularity principle leads to a convergence guarantee in nonconvex optimization and the convergence rate depends on a regularization parameter. Empirically, we find that DNN trainings that apply the training techniques achieve a fast convergence and obey the regularity principle with a large regularization parameter, implying that the model updates are well aligned with the trajectory. On the other hand, DNN trainings without the training techniques have slow convergence and obey the regularity principle with a small regularization parameter, implying that the model updates are not well aligned with the trajectory. Therefore, different training techniques regularize the model update direction via the regularity principle to facilitate the convergence.

Kaiyi Ji, Junjie Yang, Yingbin Liang

As a popular meta-learning approach, the model-agnostic meta-learning (MAML) algorithm has been widely used due to its simplicity and effectiveness. However, the convergence of the general multi-step MAML still remains unexplored. In this paper, we develop a new theoretical framework to provide such convergence guarantee for two types of objective functions that are of interest in practice: (a) resampling case (e.g., reinforcement learning), where loss functions take the form in expectation and new data are sampled as the algorithm runs; and (b) finite-sum case (e.g., supervised learning), where loss functions take the finite-sum form with given samples. For both cases, we characterize the convergence rate and the computational complexity to attain an $ε$-accurate solution for multi-step MAML in the general nonconvex setting. In particular, our results suggest that an inner-stage stepsize needs to be chosen inversely proportional to the number $N$ of inner-stage steps in order for $N$-step MAML to have guaranteed convergence. From the technical perspective, we develop novel techniques to deal with the nested structure of the meta gradient for multi-step MAML, which can be of independent interest.

Junjie Yang, Hai Zhao

Transformer-based pre-trained language models have proven to be effective for learning contextualized language representation. However, current approaches only take advantage of the output of the encoder's final layer when fine-tuning the downstream tasks. We argue that only taking single layer's output restricts the power of pre-trained representation. Thus we deepen the representation learned by the model by fusing the hidden representation in terms of an explicit HIdden Representation Extractor (HIRE), which automatically absorbs the complementary representation with respect to the output from the final layer. Utilizing RoBERTa as the backbone encoder, our proposed improvement over the pre-trained models is shown effective on multiple natural language understanding tasks and help our model rival with the state-of-the-art models on the GLUE benchmark.

Yi Zhou, Junjie Yang, Huishuai Zhang et al.

Stochastic gradient descent (SGD) has been found to be surprisingly effective in training a variety of deep neural networks. However, there is still a lack of understanding on how and why SGD can train these complex networks towards a global minimum. In this study, we establish the convergence of SGD to a global minimum for nonconvex optimization problems that are commonly encountered in neural network training. Our argument exploits the following two important properties: 1) the training loss can achieve zero value (approximately), which has been widely observed in deep learning; 2) SGD follows a star-convex path, which is verified by various experiments in this paper. In such a context, our analysis shows that SGD, although has long been considered as a randomized algorithm, converges in an intrinsically deterministic manner to a global minimum.