On the Feature Learning in Diffusion Models

The predominant success of diffusion models in generative modeling has spurred significant interest in understanding their theoretical foundations. In this work, we propose a feature learning framework aimed at analyzing and comparing the training dynamics of diffusion models with those of traditional classification models. Our theoretical analysis demonstrates that, under identical settings, diffusion models, due to the denoising objective, are encouraged to learn more balanced and comprehensive representations of the data. In contrast, neural networks with a similar architecture trained for classification tend to prioritize learning specific patterns in the data, often focusing on easy-to-learn components. To support these theoretical insights, we conduct several experiments on both synthetic and real-world datasets, which empirically validate our findings and highlight the distinct feature learning dynamics in diffusion models compared to classification.

Beyond Surface Structure: A Causal Assessment of LLMs' Comprehension Ability

Large language models (LLMs) have shown remarkable capability in natural language tasks, yet debate persists on whether they truly comprehend deep structure (i.e., core semantics) or merely rely on surface structure (e.g., presentation format). Prior studies observe that LLMs' performance declines when intervening on surface structure, arguing their success relies on surface structure recognition. However, surface structure sensitivity does not prevent deep structure comprehension. Rigorously evaluating LLMs' capability requires analyzing both, yet deep structure is often overlooked. To this end, we assess LLMs' comprehension ability using causal mediation analysis, aiming to fully discover the capability of using both deep and surface structures. Specifically, we formulate the comprehension of deep structure as direct causal effect (DCE) and that of surface structure as indirect causal effect (ICE), respectively. To address the non-estimability of original DCE and ICE -- stemming from the infeasibility of isolating mutual influences of deep and surface structures, we develop the corresponding quantifiable surrogates, including approximated DCE (ADCE) and approximated ICE (AICE). We further apply the ADCE to evaluate a series of mainstream LLMs, showing that most of them exhibit deep structure comprehension ability, which grows along with the prediction accuracy. Comparing ADCE and AICE demonstrates closed-source LLMs rely more on deep structure, while open-source LLMs are more surface-sensitive, which decreases with model scale. Theoretically, ADCE is a bidirectional evaluation, which measures both the sufficiency and necessity of deep structure changes in causing output variations, thus offering a more comprehensive assessment than accuracy, a common evaluation in LLMs. Our work provides new insights into LLMs' deep structure comprehension and offers novel methods for LLMs evaluation.

An In-depth Investigation of Sparse Rate Reduction in Transformer-like Models

Deep neural networks have long been criticized for being black-box. To unveil the inner workings of modern neural architectures, a recent work \cite{yu2024white} proposed an information-theoretic objective function called Sparse Rate Reduction (SRR) and interpreted its unrolled optimization as a Transformer-like model called Coding Rate Reduction Transformer (CRATE). However, the focus of the study was primarily on the basic implementation, and whether this objective is optimized in practice and its causal relationship to generalization remain elusive. Going beyond this study, we derive different implementations by analyzing layer-wise behaviors of CRATE, both theoretically and empirically. To reveal the predictive power of SRR on generalization, we collect a set of model variants induced by varied implementations and hyperparameters and evaluate SRR as a complexity measure based on its correlation with generalization. Surprisingly, we find out that SRR has a positive correlation coefficient and outperforms other baseline measures, such as path-norm and sharpness-based ones. Furthermore, we show that generalization can be improved using SRR as regularization on benchmark image classification datasets. We hope this paper can shed light on leveraging SRR to design principled models and study their generalization ability.

How Does Critical Batch Size Scale in Pre-training?

Training large-scale models under given resources requires careful design of parallelism strategies. In particular, the efficiency notion of critical batch size, concerning the compromise between time and compute, marks the threshold beyond which greater data parallelism leads to diminishing returns. To operationalize it, we propose a measure of CBS and pre-train a series of auto-regressive language models, ranging from 85 million to 1.2 billion parameters, on the C4 dataset. Through extensive hyper-parameter sweeps and careful control on factors such as batch size, momentum, and learning rate along with its scheduling, we systematically investigate the impact of scale on CBS. Then we fit scaling laws with respect to model and data sizes to decouple their effects. Overall, our results demonstrate that CBS scales primarily with data size rather than model size, a finding we justify theoretically through the analysis of infinite-width limits of neural networks and infinite-dimensional least squares regression. Of independent interest, we highlight the importance of common hyper-parameter choices and strategies for studying large-scale pre-training beyond fixed training durations.

Initialization Matters: On the Benign Overfitting of Two-Layer ReLU CNN with Fully Trainable Layers

Benign overfitting refers to how over-parameterized neural networks can fit training data perfectly and generalize well to unseen data. While this has been widely investigated theoretically, existing works are limited to two-layer networks with fixed output layers, where only the hidden weights are trained. We extend the analysis to two-layer ReLU convolutional neural networks (CNNs) with fully trainable layers, which is closer to the practice. Our results show that the initialization scaling of the output layer is crucial to the training dynamics: large scales make the model training behave similarly to that with the fixed output, the hidden layer grows rapidly while the output layer remains largely unchanged; in contrast, small scales result in more complex layer interactions, the hidden layer initially grows to a specific ratio relative to the output layer, after which both layers jointly grow and maintain that ratio throughout training. Furthermore, in both settings, we provide nearly matching upper and lower bounds on the test errors, identifying the sharp conditions on the initialization scaling and signal-to-noise ratio (SNR) in which the benign overfitting can be achieved or not. Numerical experiments back up the theoretical results.

How Transformers Utilize Multi-Head Attention in In-Context Learning? A Case Study on Sparse Linear Regression

Despite the remarkable success of transformer-based models in various real-world tasks, their underlying mechanisms remain poorly understood. Recent studies have suggested that transformers can implement gradient descent as an in-context learner for linear regression problems and have developed various theoretical analyses accordingly. However, these works mostly focus on the expressive power of transformers by designing specific parameter constructions, lacking a comprehensive understanding of their inherent working mechanisms post-training. In this study, we consider a sparse linear regression problem and investigate how a trained multi-head transformer performs in-context learning. We experimentally discover that the utilization of multi-heads exhibits different patterns across layers: multiple heads are utilized and essential in the first layer, while usually only a single head is sufficient for subsequent layers. We provide a theoretical explanation for this observation: the first layer preprocesses the context data, and the following layers execute simple optimization steps based on the preprocessed context. Moreover, we demonstrate that such a preprocess-then-optimize algorithm can significantly outperform naive gradient descent and ridge regression algorithms. Further experimental results support our explanations. Our findings offer insights into the benefits of multi-head attention and contribute to understanding the more intricate mechanisms hidden within trained transformers.

Extracting Training Data from Unconditional Diffusion Models

As diffusion probabilistic models (DPMs) are being employed as mainstream models for generative artificial intelligence (AI), the study of their memorization of the raw training data has attracted growing attention. Existing works in this direction aim to establish an understanding of whether or to what extent DPMs learn by memorization. Such an understanding is crucial for identifying potential risks of data leakage and copyright infringement in diffusion models and, more importantly, for more controllable generation and trustworthy application of Artificial Intelligence Generated Content (AIGC). While previous works have made important observations of when DPMs are prone to memorization, these findings are mostly empirical, and the developed data extraction methods only work for conditional diffusion models. In this work, we aim to establish a theoretical understanding of memorization in DPMs with 1) a memorization metric for theoretical analysis, 2) an analysis of conditional memorization with informative and random labels, and 3) two better evaluation metrics for measuring memorization. Based on the theoretical analysis, we further propose a novel data extraction method called \textbf{Surrogate condItional Data Extraction (SIDE)} that leverages a classifier trained on generated data as a surrogate condition to extract training data directly from unconditional diffusion models. Our empirical results demonstrate that SIDE can extract training data from diffusion models where previous methods fail, and it is on average over 50\% more effective across different scales of the CelebA dataset.

Explainable Bayesian Recurrent Neural Smoother to Capture Global State Evolutionary Correlations

Through integrating the evolutionary correlations across global states in the bidirectional recursion, an explainable Bayesian recurrent neural smoother (EBRNS) is proposed for offline data-assisted fixed-interval state smoothing. At first, the proposed model, containing global states in the evolutionary interval, is transformed into an equivalent model with bidirectional memory. This transformation incorporates crucial global state information with support for bi-directional recursive computation. For the transformed model, the joint state-memory-trend Bayesian filtering and smoothing frameworks are derived by introducing the bidirectional memory iteration mechanism and offline data into Bayesian estimation theory. The derived frameworks are implemented using the Gaussian approximation to ensure analytical properties and computational efficiency. Finally, the neural network modules within EBRNS and its two-stage training scheme are designed. Unlike most existing approaches that artificially combine deep learning and model-based estimation, the bidirectional recursion and internal gated structures of EBRNS are naturally derived from Bayesian estimation theory, explainably integrating prior model knowledge, online measurement, and offline data. Experiments on representative real-world datasets demonstrate that the high smoothing accuracy of EBRNS is accompanied by data efficiency and a lightweight parameter scale.

The Implicit Bias of Adam on Separable Data

Adam has become one of the most favored optimizers in deep learning problems. Despite its success in practice, numerous mysteries persist regarding its theoretical understanding. In this paper, we study the implicit bias of Adam in linear logistic regression. Specifically, we show that when the training data are linearly separable, Adam converges towards a linear classifier that achieves the maximum $\ell_\infty$-margin. Notably, for a general class of diminishing learning rates, this convergence occurs within polynomial time. Our result shed light on the difference between Adam and (stochastic) gradient descent from a theoretical perspective.

Self-Control of LLM Behaviors by Compressing Suffix Gradient into Prefix Controller

We propose Self-Control, a novel method utilizing suffix gradients to control the behavior of large language models (LLMs) without explicit human annotations. Given a guideline expressed in suffix string and the model's self-assessment of adherence, Self-Control computes the gradient of this self-judgment concerning the model's hidden states, directly influencing the auto-regressive generation process towards desired behaviors. To enhance efficiency, we introduce Self-Control_{prefix}, a compact module that encapsulates the learned representations from suffix gradients into a Prefix Controller, facilitating inference-time control for various LLM behaviors. Our experiments demonstrate Self-Control's efficacy across multiple domains, including emotional modulation, ensuring harmlessness, and enhancing complex reasoning. Especially, Self-Control_{prefix} enables a plug-and-play control and jointly controls multiple attributes, improving model outputs without altering model parameters or increasing inference-time costs.