The training process of ReLU neural networks often exhibits complicated nonlinear phenomena. The nonlinearity of models and non-convexity of loss pose significant challenges for theoretical analysis. Therefore, most previous theoretical works on the optimization dynamics of neural networks focus either on local analysis (like the end of training) or approximate linear models (like Neural Tangent Kernel). In this work, we conduct a complete theoretical characterization of the training process of a two-layer ReLU network trained by Gradient Flow on a linearly separable data. In this specific setting, our analysis captures the whole optimization process starting from random initialization to final convergence. Despite the relatively simple model and data that we studied, we reveal four different phases from the whole training process showing a general simplifying-to-complicating learning trend. Specific nonlinear behaviors can also be precisely identified and captured theoretically, such as initial condensation, saddle-to-plateau dynamics, plateau escape, changes of activation patterns, learning with increasing complexity, etc.

Prompt learning has emerged as an effective and data-efficient technique in large Vision-Language Models (VLMs). However, when adapting VLMs to specialized domains such as remote sensing and medical imaging, domain prompt learning remains underexplored. While large-scale domain-specific foundation models can help tackle this challenge, their concentration on a single vision level makes it challenging to prompt both vision and language modalities. To overcome this, we propose to leverage domain-specific knowledge from domain-specific foundation models to transfer the robust recognition ability of VLMs from generalized to specialized domains, using quaternion networks. Specifically, the proposed method involves using domain-specific vision features from domain-specific foundation models to guide the transformation of generalized contextual embeddings from the language branch into a specialized space within the quaternion networks. Moreover, we present a hierarchical approach that generates vision prompt features by analyzing intermodal relationships between hierarchical language prompt features and domain-specific vision features. In this way, quaternion networks can effectively mine the intermodal relationships in the specific domain, facilitating domain-specific vision-language contrastive learning. Extensive experiments on domain-specific datasets show that our proposed method achieves new state-of-the-art results in prompt learning.

The gold standard for causal model evaluation involves comparing model predictions with true effects estimated from randomized controlled trials (RCT). However, RCTs are not always feasible or ethical to perform. In contrast, conditionally randomized experiments based on inverse probability weighting (IPW) offer a more realistic approach but may suffer from high estimation variance. To tackle this challenge and enhance causal model evaluation in real-world conditional randomization settings, we introduce a novel low-variance estimator for causal error, dubbed as the pairs estimator. By applying the same IPW estimator to both the model and true experimental effects, our estimator effectively cancels out the variance due to IPW and achieves a smaller asymptotic variance. Empirical studies demonstrate the improved of our estimator, highlighting its potential on achieving near-RCT performance. Our method offers a simple yet powerful solution to evaluate causal inference models in conditional randomization settings without complicated modification of the IPW estimator itself, paving the way for more robust and reliable model assessments.

Recent advances in artificial intelligence have created a paradigm shift in which models are trained on large amounts of data and can be adapted to different tasks, dubbed foundation models (Bommasani et al., 2021). These models, which often employ self-supervision, can extract valuable knowledge from various types of data, including natural language (Devlin et al., 2018; Brown et al., 2020), images (Radford et al., 2021), and biological sequencing counts (Theodoris et al., 2023). This acquired knowledge allows the model to generalize when asked to perform tasks in novel scenarios. With vast amounts of data becoming increasingly available from diverse sources, such models are of interest to leverage information that can be learned in order to build more intelligent systems (Bubeck et al., 2023). A critical aspect of intelligent systems is the ability to reason about cause-and-effect relationships (Zhang et al., 2023), which is vital to making informed decisions across various domains, including healthcare, economics, and statistics (Kube et al., 2019; Geffner et al., 2022; Zhang et al., 2022). Relying solely on correlation-based models (Harrison and March, 1984) can lead to misleading conclusions, as they do not account for the underlying causal mechanisms. This limitation is also observed in the realm of foundation models (Bubeck et al., 2023; Mahowald et al., 2023; Wolfram, 2023).

Generalization error bounds for deep neural networks trained by stochastic gradient descent (SGD) are derived by combining a dynamical control of an appropriate parameter norm and the Rademacher complexity estimate based on parameter norms. The bounds explicitly depend on the loss along the training trajectory, and work for a wide range of network architectures including multilayer perceptron (MLP) and convolutional neural networks (CNN). Compared with other algorithm-depending generalization estimates such as uniform stability-based bounds, our bounds do not require $L$-smoothness of the nonconvex loss function, and apply directly to SGD instead of Stochastic Langevin gradient descent (SGLD). Numerical results show that our bounds are non-vacuous and robust with the change of optimizer and network hyperparameters.

The convergence of GD and SGD when training mildly parameterized neural networks starting from random initialization is studied. For a broad range of models and loss functions, including the most commonly used square loss and cross entropy loss, we prove an ``early stage convergence'' result. We show that the loss is decreased by a significant amount in the early stage of the training, and this decrease is fast. Furthurmore, for exponential type loss functions, and under some assumptions on the training data, we show global convergence of GD. Instead of relying on extreme over-parameterization, our study is based on a microscopic analysis of the activation patterns for the neurons, which helps us derive more powerful lower bounds for the gradient. The results on activation patterns, which we call ``neuron partition'', help build intuitions for understanding the behavior of neural networks' training dynamics, and may be of independent interest.

We assess the ability of large language models (LLMs) to answer causal questions by analyzing their strengths and weaknesses against three types of causal question. We believe that current LLMs can answer causal questions with existing causal knowledge as combined domain experts. However, they are not yet able to provide satisfactory answers for discovering new knowledge or for high-stakes decision-making tasks with high precision. We discuss possible future directions and opportunities, such as enabling explicit and implicit causal modules as well as deep causal-aware LLMs. These will not only enable LLMs to answer many different types of causal questions for greater impact but also enable LLMs to be more trustworthy and efficient in general.

Latent confounding has been a long-standing obstacle for causal reasoning from observational data. One popular approach is to model the data using acyclic directed mixed graphs (ADMGs), which describe ancestral relations between variables using directed and bidirected edges. However, existing methods using ADMGs are based on either linear functional assumptions or a discrete search that is complicated to use and lacks computational tractability for large datasets. In this work, we further extend the existing body of work and develop a novel gradient-based approach to learning an ADMG with non-linear functional relations from observational data. We first show that the presence of latent confounding is identifiable under the assumptions of bow-free ADMGs with non-linear additive noise models. With this insight, we propose a novel neural causal model based on autoregressive flows for ADMG learning. This not only enables us to determine complex causal structural relationships behind the data in the presence of latent confounding, but also estimate their functional relationships (hence treatment effects) simultaneously. We further validate our approach via experiments on both synthetic and real-world datasets, and demonstrate the competitive performance against relevant baselines.

In this work, we explore the maximum-margin bias of quasi-homogeneous neural networks trained with gradient flow on an exponential loss and past a point of separability. We introduce the class of quasi-homogeneous models, which is expressive enough to describe nearly all neural networks with homogeneous activations, even those with biases, residual connections, and normalization layers, while structured enough to enable geometric analysis of its gradient dynamics. Using this analysis, we generalize the existing results of maximum-margin bias for homogeneous networks to this richer class of models. We find that gradient flow implicitly favors a subset of the parameters, unlike in the case of a homogeneous model where all parameters are treated equally. We demonstrate through simple examples how this strong favoritism toward minimizing an asymmetric norm can degrade the robustness of quasi-homogeneous models. On the other hand, we conjecture that this norm-minimization discards, when possible, unnecessary higher-order parameters, reducing the model to a sparser parameterization. Lastly, by applying our theorem to sufficiently expressive neural networks with normalization layers, we reveal a universal mechanism behind the empirical phenomenon of Neural Collapse.

ChatGPT[3] has demonstrated exceptional proficiency in natural language conversation, e.g., it can answer a wide range of questions while no previous large language models can. Thus, we would like to push its limit and explore its ability to answer causal discovery questions by using a medical benchmark [5] in causal discovery. Causal discovery aims to uncover the underlying unknown causal relationships based purely on observational data[2]. In contrast, applying ChatGPT to answer the questions about causal relationships is fundamentally different. With the current version of ChatGPT, we can only use the names (meta information) instead of observational data of variables to answer causal questions. The answers to the causal questions given by ChatGPT are based on a trained large language model, which can be viewed as an approximation for existing knowledge in the training natural language data.