In this paper, we introduce the Dependent Noise-based Inaccurate Label Distribution Learning (DN-ILDL) framework to tackle the challenges posed by noise in label distribution learning, which arise from dependencies on instances and labels. We start by modeling the inaccurate label distribution matrix as a combination of the true label distribution and a noise matrix influenced by specific instances and labels. To address this, we develop a linear mapping from instances to their true label distributions, incorporating label correlations, and decompose the noise matrix using feature and label representations, applying group sparsity constraints to accurately capture the noise. Furthermore, we employ graph regularization to align the topological structures of the input and output spaces, ensuring accurate reconstruction of the true label distribution matrix. Utilizing the Alternating Direction Method of Multipliers (ADMM) for efficient optimization, we validate our method's capability to recover true labels accurately and establish a generalization error bound. Extensive experiments demonstrate that DN-ILDL effectively addresses the ILDL problem and outperforms existing LDL methods.

Offline Black-Box Optimization (BBO) aims at optimizing a black-box function using the knowledge from a pre-collected offline dataset of function values and corresponding input designs. However, the high-dimensional and highly-multimodal input design space of black-box function pose inherent challenges for most existing methods that model and operate directly upon input designs. These issues include but are not limited to high sample complexity, which relates to inaccurate approximation of black-box function; and insufficient coverage and exploration of input design modes, which leads to suboptimal proposal of new input designs. In this work, we consider finding a latent space that serves as a compressed yet accurate representation of the design-value joint space, enabling effective latent exploration of high-value input design modes. To this end, we formulate an learnable energy-based latent space, and propose Noise-intensified Telescoping density-Ratio Estimation (NTRE) scheme for variational learning of an accurate latent space model without costly Markov Chain Monte Carlo. The optimization process is then exploration of high-value designs guided by the learned energy-based model in the latent space, formulated as gradient-based sampling from a latent-variable-parameterized inverse model. We show that our particular parameterization encourages expanded exploration around high-value design modes, motivated by inversion thinking of a fundamental result of conditional covariance matrix typically used for variance reduction. We observe that our method, backed by an accurately learned informative latent space and an expanding-exploration model design, yields significant improvements over strong previous methods on both synthetic and real world datasets such as the design-bench suite.

Desirable random graph models (RGMs) should (i) be tractable so that we can compute and control graph statistics, and (ii) generate realistic structures such as high clustering (i.e., high subgraph densities). A popular category of RGMs (e.g., Erdos-Renyi and stochastic Kronecker) outputs edge probabilities, and we need to realize (i.e., sample from) the edge probabilities to generate graphs. Typically, each edge (in)existence is assumed to be determined independently. However, with edge independency, RGMs theoretically cannot produce high subgraph densities unless they "replicate" input graphs. In this work, we explore realization beyond edge independence that can produce more realistic structures while ensuring high tractability. Specifically, we propose edge-dependent realization schemes called binding and derive closed-form tractability results on subgraph (e.g., triangle) densities in graphs generated with binding. We propose algorithms for graph generation with binding and parameter fitting of binding. We empirically validate that binding exhibits high tractability and generates realistic graphs with high clustering, significantly improving upon existing RGMs assuming edge independency.

Past work has established scaling laws that predict the performance of a neural language model (LM) as a function of its parameter count and the number of tokens it's trained on, enabling optimal allocation of a fixed compute budget. Are these scaling laws agnostic to training data as some prior work suggests? We generate training datasets of varying complexities by modulating the syntactic properties of a PCFG, finding that 1) scaling laws are sensitive to differences in data complexity and that 2) gzip, a compression algorithm, is an effective predictor of how data complexity impacts scaling properties. We propose a new data-dependent scaling law for LM's that accounts for the training data's gzip-compressibility; its compute-optimal frontier increases in dataset size preference (over parameter count preference) as training data becomes harder to compress.

Graph neural networks (GNNs) have exhibited prominent performance in learning graph-structured data. Considering node classification task, based on the i.i.d assumption among node labels, the traditional supervised learning simply sums up cross-entropy losses of the independent training nodes and applies the average loss to optimize GNNs' weights. But different from other data formats, the nodes are naturally connected. It is found that the independent distribution modeling of node labels restricts GNNs' capability to generalize over the entire graph and defend adversarial attacks. In this work, we propose a new framework, termed joint-cluster supervised learning, to model the joint distribution of each node with its corresponding cluster. We learn the joint distribution of node and cluster labels conditioned on their representations, and train GNNs with the obtained joint loss. In this way, the data-label reference signals extracted from the local cluster explicitly strengthen the discrimination ability on the target node. The extensive experiments demonstrate that our joint-cluster supervised learning can effectively bolster GNNs' node classification accuracy. Furthermore, being benefited from the reference signals which may be free from spiteful interference, our learning paradigm significantly protects the node classification from being affected by the adversarial attack.

Advancements in language foundation models have primarily fueled the recent surge in artificial intelligence. In contrast, generative learning of non-textual modalities, especially videos, significantly trails behind language modeling. This thesis chronicles our endeavor to build multi-task models for generating videos and other modalities under diverse conditions, as well as for understanding and compression applications. Given the high dimensionality of visual data, we pursue concise and accurate latent representations. Our video-native spatial-temporal tokenizers preserve high fidelity. We unveil a novel approach to mapping bidirectionally between visual observation and interpretable lexical terms. Furthermore, our scalable visual token representation proves beneficial across generation, compression, and understanding tasks. This achievement marks the first instances of language models surpassing diffusion models in visual synthesis and a video tokenizer outperforming industry-standard codecs. Within these multi-modal latent spaces, we study the design of multi-task generative models. Our masked multi-task transformer excels at the quality, efficiency, and flexibility of video generation. We enable a frozen language model, trained solely on text, to generate visual content. Finally, we build a scalable generative multi-modal transformer trained from scratch, enabling the generation of videos containing high-fidelity motion with the corresponding audio given diverse conditions. Throughout the course, we have shown the effectiveness of integrating multiple tasks, crafting high-fidelity latent representation, and generating multiple modalities. This work suggests intriguing potential for future exploration in generating non-textual data and enabling real-time, interactive experiences across various media forms.

In recent years, Transformers have become the de-facto architecture for long-term sequence forecasting (LTSF), but faces challenges such as quadratic complexity and permutation invariant bias. A recent model, Mamba, based on selective state space models (SSMs), has emerged as a competitive alternative to Transformer, offering comparable performance with higher throughput and linear complexity related to sequence length. In this study, we analyze the limitations of current Mamba in LTSF and propose four targeted improvements, leading to MambaTS. We first introduce variable scan along time to arrange the historical information of all the variables together. We suggest that causal convolution in Mamba is not necessary for LTSF and propose the Temporal Mamba Block (TMB). We further incorporate a dropout mechanism for selective parameters of TMB to mitigate model overfitting. Moreover, we tackle the issue of variable scan order sensitivity by introducing variable permutation training. We further propose variable-aware scan along time to dynamically discover variable relationships during training and decode the optimal variable scan order by solving the shortest path visiting all nodes problem during inference. Extensive experiments conducted on eight public datasets demonstrate that MambaTS achieves new state-of-the-art performance.

Neural networks are widespread due to their powerful performance. However, they degrade in the presence of noisy labels at training time. Inspired by the setting of learning with expert advice, where multiplicative weight (MW) updates were recently shown to be robust to moderate data corruptions in expert advice, we propose to use MW for reweighting examples during neural networks optimization. We theoretically establish the convergence of our method when used with gradient descent and prove its advantages in 1d cases. We then validate our findings empirically for the general case by showing that MW improves the accuracy of neural networks in the presence of label noise on CIFAR-10, CIFAR-100 and Clothing1M. We also show the impact of our approach on adversarial robustness.

Stochastic gradient descent (SGD) exhibits strong algorithmic regularization effects in practice and plays an important role in the generalization of modern machine learning. However, prior research has revealed instances where the generalization performance of SGD is worse than ridge regression due to uneven optimization along different dimensions. Preconditioning offers a natural solution to this issue by rebalancing optimization across different directions. Yet, the extent to which preconditioning can enhance the generalization performance of SGD and whether it can bridge the existing gap with ridge regression remains uncertain. In this paper, we study the generalization performance of SGD with preconditioning for the least squared problem. We make a comprehensive comparison between preconditioned SGD and (standard \& preconditioned) ridge regression. Our study makes several key contributions toward understanding and improving SGD with preconditioning. First, we establish excess risk bounds (generalization performance) for preconditioned SGD and ridge regression under an arbitrary preconditions matrix. Second, leveraging the excessive risk characterization of preconditioned SGD and ridge regression, we show that (through construction) there exists a simple preconditioned matrix that can make SGD comparable to (standard \& preconditioned) ridge regression. Finally, we show that our proposed preconditioning matrix is straightforward enough to allow robust estimation from finite samples while maintaining a theoretical improvement. Our empirical results align with our theoretical findings, collectively showcasing the enhanced regularization effect of preconditioned SGD.

Effectively representing heterogeneous tabular datasets for meta-learning remains an open problem. Previous approaches rely on predefined meta-features, for example, statistical measures or landmarkers. Encoder-based models, such as Dataset2Vec, allow us to extract significant meta-features automatically without human intervention. This research introduces a novel encoder-based representation of tabular datasets implemented within the liltab package available on GitHub https://github.com/azoz01/liltab. Our package is based on an established model for heterogeneous tabular data proposed in [Tomoharu Iwata and Atsutoshi Kumagai. Meta-learning from Tasks with Heterogeneous Attribute Spaces. In Advances in Neural Information Processing Systems, 2020]. The proposed approach employs a different model for encoding feature relationships, generating alternative representations compared to existing methods like Dataset2Vec. Both of them leverage the fundamental assumption of dataset similarity learning. In this work, we evaluate Dataset2Vec and liltab on two common meta-tasks -- representing entire datasets and hyperparameter optimization warm-start. However, validation on an independent metaMIMIC dataset highlights the nuanced challenges in representation learning. We show that general representations may not suffice for some meta-tasks where requirements are not explicitly considered during extraction.