Deep representations have shown promising performance when transferred to downstream tasks in a black-box manner. Yet, their inherent lack of interpretability remains a significant challenge, as these features are often opaque to human understanding. In this paper, we propose Non-negative Contrastive Learning (NCL), a renaissance of Non-negative Matrix Factorization (NMF) aimed at deriving interpretable features. The power of NCL lies in its enforcement of non-negativity constraints on features, reminiscent of NMF's capability to extract features that align closely with sample clusters. NCL not only aligns mathematically well with an NMF objective but also preserves NMF's interpretability attributes, resulting in a more sparse and disentangled representation compared to standard contrastive learning (CL). Theoretically, we establish guarantees on the identifiability and downstream generalization of NCL. Empirically, we show that these advantages enable NCL to outperform CL significantly on feature disentanglement, feature selection, as well as downstream classification tasks. At last, we show that NCL can be easily extended to other learning scenarios and benefit supervised learning as well. Code is available at https://github.com/PKU-ML/non_neg.
Contrastive Learning (CL) has emerged as one of the most successful paradigms for unsupervised visual representation learning, yet it often depends on intensive manual data augmentations. With the rise of generative models, especially diffusion models, the ability to generate realistic images close to the real data distribution has been well recognized. These generated high-equality images have been successfully applied to enhance contrastive representation learning, a technique termed ``data inflation''. However, we find that the generated data (even from a good diffusion model like DDPM) may sometimes even harm contrastive learning. We investigate the causes behind this failure from the perspective of both data inflation and data augmentation. For the first time, we reveal the complementary roles that stronger data inflation should be accompanied by weaker augmentations, and vice versa. We also provide rigorous theoretical explanations for these phenomena via deriving its generalization bounds under data inflation. Drawing from these insights, we propose Adaptive Inflation (AdaInf), a purely data-centric strategy without introducing any extra computation cost. On benchmark datasets, AdaInf can bring significant improvements for various contrastive learning methods. Notably, without using external data, AdaInf obtains 94.70% linear accuracy on CIFAR-10 with SimCLR, setting a new record that surpasses many sophisticated methods. Code is available at https://github.com/PKU-ML/adainf.
We prove that training neural networks on 1-D data is equivalent to solving a convex Lasso problem with a fixed, explicitly defined dictionary matrix of features. The specific dictionary depends on the activation and depth. We consider 2-layer networks with piecewise linear activations, deep narrow ReLU networks with up to 4 layers, and rectangular and tree networks with sign activation and arbitrary depth. Interestingly in ReLU networks, a fourth layer creates features that represent reflections of training data about themselves. The Lasso representation sheds insight to globally optimal networks and the solution landscape.
Adversarial Training (AT), which adversarially perturb the input samples during training, has been acknowledged as one of the most effective defenses against adversarial attacks, yet suffers from a fundamental tradeoff that inevitably decreases clean accuracy. Instead of perturbing the samples, Sharpness-Aware Minimization (SAM) perturbs the model weights during training to find a more flat loss landscape and improve generalization. However, as SAM is designed for better clean accuracy, its effectiveness in enhancing adversarial robustness remains unexplored. In this work, considering the duality between SAM and AT, we investigate the adversarial robustness derived from SAM. Intriguingly, we find that using SAM alone can improve adversarial robustness. To understand this unexpected property of SAM, we first provide empirical and theoretical insights into how SAM can implicitly learn more robust features, and conduct comprehensive experiments to show that SAM can improve adversarial robustness notably without sacrificing any clean accuracy, shedding light on the potential of SAM to be a substitute for AT when accuracy comes at a higher priority. Code is available at https://github.com/weizeming/SAM_AT.
Learning-outcome prediction (LOP) is a long-standing and critical problem in educational routes. Many studies have contributed to developing effective models while often suffering from data shortage and low generalization to various institutions due to the privacy-protection issue. To this end, this study proposes a distributed grade prediction model, dubbed FecMap, by exploiting the federated learning (FL) framework that preserves the private data of local clients and communicates with others through a global generalized model. FecMap considers local subspace learning (LSL), which explicitly learns the local features against the global features, and multi-layer privacy protection (MPP), which hierarchically protects the private features, including model-shareable features and not-allowably shared features, to achieve client-specific classifiers of high performance on LOP per institution. FecMap is then achieved in an iteration manner with all datasets distributed on clients by training a local neural network composed of a global part, a local part, and a classification head in clients and averaging the global parts from clients on the server. To evaluate the FecMap model, we collected three higher-educational datasets of student academic records from engineering majors. Experiment results manifest that FecMap benefits from the proposed LSL and MPP and achieves steady performance on the task of LOP, compared with the state-of-the-art models. This study makes a fresh attempt at the use of federated learning in the learning-analytical task, potentially paving the way to facilitating personalized education with privacy protection.
Researchers have recently found that Self-Supervised Learning (SSL) is vulnerable to backdoor attacks. The attacker can embed hidden SSL backdoors via a few poisoned examples in the training dataset and maliciously manipulate the behavior of downstream models. To defend against SSL backdoor attacks, a feasible route is to detect and remove the poisonous samples in the training set. However, the existing SSL backdoor defense method fails to detect the poisonous samples precisely. In this paper, we propose to erase the SSL backdoor by cluster activation masking and propose a novel PoisonCAM method. After obtaining the threat model trained on the poisoned dataset, our method can precisely detect poisonous samples based on the assumption that masking the backdoor trigger can effectively change the activation of a downstream clustering model. In experiments, our PoisonCAM achieves 96% accuracy for backdoor trigger detection compared to 3% of the state-of-the-art method on poisoned ImageNet-100. Moreover, our proposed PoisonCAM significantly improves the performance of the trained SSL model under backdoor attacks compared to the state-of-the-art method. Our code will be available at https://github.com/LivXue/PoisonCAM.
Trajectory prediction is a challenging task that aims to predict the future trajectory of vehicles or pedestrians over a short time horizon based on their historical positions. The main reason is that the trajectory is a kind of complex data, including spatial and temporal information, which is crucial for accurate prediction. Intuitively, the more information the model can capture, the more precise the future trajectory can be predicted. However, previous works based on deep learning methods processed spatial and temporal information separately, leading to inadequate spatial information capture, which means they failed to capture the complete spatial information. Therefore, it is of significance to capture information more fully and effectively on vehicle interactions. In this study, we introduced an integrated 3D graph that incorporates both spatial and temporal edges. Based on this, we proposed the integrated 3D graph, which considers the cross-time interaction information. In specific, we design a Spatial-Temporal Fusion (STF) model including Multi-layer perceptions (MLP) and Graph Attention (GAT) to capture the spatial and temporal information historical trajectories simultaneously on the 3D graph. Our experiment on the ApolloScape Trajectory Datasets shows that the proposed STF outperforms several baseline methods, especially on the long-time-horizon trajectory prediction.
The versatility of multimodal deep learning holds tremendous promise for advancing scientific research and practical applications. As this field continues to evolve, the collective power of cross-modal analysis promises to drive transformative innovations, leading us to new frontiers in chemical understanding and discovery. Hence, we introduce Asymmetric Contrastive Multimodal Learning (ACML) as a novel approach tailored for molecules, showcasing its potential to advance the field of chemistry. ACML harnesses the power of effective asymmetric contrastive learning to seamlessly transfer information from various chemical modalities to molecular graph representations. By combining pre-trained chemical unimodal encoders and a shallow-designed graph encoder, ACML facilitates the assimilation of coordinated chemical semantics from different modalities, leading to comprehensive representation learning with efficient training. This innovative framework enhances the interpretability of learned representations and bolsters the expressive power of graph neural networks. Through practical tasks such as isomer discrimination and uncovering crucial chemical properties for drug discovery, ACML exhibits its capability to revolutionize chemical research and applications, providing a deeper understanding of chemical semantics of different modalities.
We investigate the complexity of training a two-layer ReLU neural network with weight decay regularization. Previous research has shown that the optimal solution of this problem can be found by solving a standard cone-constrained convex program. Using this convex formulation, we prove that the hardness of approximation of ReLU networks not only mirrors the complexity of the Max-Cut problem but also, in certain special cases, exactly corresponds to it. In particular, when $\epsilon\leq\sqrt{84/83}-1\approx 0.006$, we show that it is NP-hard to find an approximate global optimizer of the ReLU network objective with relative error $\epsilon$ with respect to the objective value. Moreover, we develop a randomized algorithm which mirrors the Goemans-Williamson rounding of semidefinite Max-Cut relaxations. To provide polynomial-time approximations, we classify training datasets into three categories: (i) For orthogonal separable datasets, a precise solution can be obtained in polynomial-time. (ii) When there is a negative correlation between samples of different classes, we give a polynomial-time approximation with relative error $\sqrt{\pi/2}-1\approx 0.253$. (iii) For general datasets, the degree to which the problem can be approximated in polynomial-time is governed by a geometric factor that controls the diameter of two zonotopes intrinsic to the dataset. To our knowledge, these results present the first polynomial-time approximation guarantees along with first hardness of approximation results for regularized ReLU networks.
With the prosperity of contrastive learning for visual representation learning (VCL), it is also adapted to the graph domain and yields promising performance. However, through a systematic study of various graph contrastive learning (GCL) methods, we observe that some common phenomena among existing GCL methods that are quite different from the original VCL methods, including 1) positive samples are not a must for GCL; 2) negative samples are not necessary for graph classification, neither for node classification when adopting specific normalization modules; 3) data augmentations have much less influence on GCL, as simple domain-agnostic augmentations (e.g., Gaussian noise) can also attain fairly good performance. By uncovering how the implicit inductive bias of GNNs works in contrastive learning, we theoretically provide insights into the above intriguing properties of GCL. Rather than directly porting existing VCL methods to GCL, we advocate for more attention toward the unique architecture of graph learning and consider its implicit influence when designing GCL methods. Code is available at https: //github.com/PKU-ML/ArchitectureMattersGCL.