The goal of the Inverse reinforcement learning (IRL) task is to identify the underlying reward function and the corresponding optimal policy from a set of expert demonstrations. While most IRL algorithms' theoretical guarantees rely on a linear reward structure, we aim to extend the theoretical understanding of IRL to scenarios where the reward function is parameterized by neural networks. Meanwhile, conventional IRL algorithms usually adopt a nested structure, leading to computational inefficiency, especially in high-dimensional settings. To address this problem, we propose the first two-timescale single-loop IRL algorithm under neural network parameterized reward and provide a non-asymptotic convergence analysis under overparameterization. Although prior optimality results for linear rewards do not apply, we show that our algorithm can identify the globally optimal reward and policy under certain neural network structures. This is the first IRL algorithm with a non-asymptotic convergence guarantee that provably achieves global optimality in neural network settings.

There are many approaches for training decision trees. This work introduces a novel gradient-based method for constructing decision trees that optimize arbitrary differentiable loss functions, overcoming the limitations of heuristic splitting rules. Unlike traditional approaches that rely on heuristic splitting rules, the proposed method refines predictions using the first and second derivatives of the loss function, enabling the optimization of complex tasks such as classification, regression, and survival analysis. We demonstrate the method's applicability to classification, regression, and survival analysis tasks, including those with censored data. Numerical experiments on both real and synthetic datasets compare the proposed method with traditional decision tree algorithms, such as CART, Extremely Randomized Trees, and SurvTree. The implementation of the method is publicly available, providing a practical tool for researchers and practitioners. This work advances the field of decision tree-based modeling, offering a more flexible and accurate approach for handling structured data and complex tasks. By leveraging gradient-based optimization, the proposed method bridges the gap between traditional decision trees and modern machine learning techniques, paving the way for further innovations in interpretable and high-performing models.

Causal discovery (CD) plays a pivotal role in numerous scientific fields by clarifying the causal relationships that underlie phenomena observed in diverse disciplines. Despite significant advancements in CD algorithms that enhance bias and fairness analyses in machine learning, their application faces challenges due to the high computational demands and complexities of large-scale data. This paper introduces a framework that leverages Large Language Models (LLMs) for CD, utilizing a metadata-based approach akin to the reasoning processes of human experts. By shifting from pairwise queries to a more scalable breadth-first search (BFS) strategy, the number of required queries is reduced from quadratic to linear in terms of variable count, thereby addressing scalability concerns inherent in previous approaches. This method utilizes an Active Learning (AL) and a Dynamic Scoring Mechanism that prioritizes queries based on their potential information gain, combining mutual information, partial correlation, and LLM confidence scores to refine the causal graph more efficiently and accurately. This BFS query strategy reduces the required number of queries significantly, thereby addressing scalability concerns inherent in previous approaches. This study provides a more scalable and efficient solution for leveraging LLMs in fairness-driven CD, highlighting the effects of the different parameters on performance. We perform fairness analyses on the inferred causal graphs, identifying direct and indirect effects of sensitive attributes on outcomes. A comparison of these analyses against those from graphs produced by baseline methods highlights the importance of accurate causal graph construction in understanding bias and ensuring fairness in machine learning systems.

The outcome of Large Language Model (LLM) pre-training strongly depends on weight initialization and variance control strategies. Although the importance of initial variance control has been well documented in neural networks in general, the literature on initialization and management of its growth during LLM pre-training, specifically, is somewhat sparse. In this paper, we introduce the Layer Index Rescaling (LIR) weight initialization scheme, and the Target Variance Rescaling (TVR) variance control strategy. Experiments on a 1B parameter LLaMA model demonstrate that better variance management using these techniques yields substantial improvements in downstream task performance (up to 4.6% on common pre-training benchmarks) and reduces extreme activation values, thus mitigating challenges associated with quantization and low-precision training.

Knowledge distillation is a technique used to train a small student network using the output generated by a large teacher network, and has many empirical advantages~\citep{Hinton2015DistillingTK}. While the standard one-shot approach to distillation only uses the output of the final teacher network, recent work~\citep{panigrahi2024progressive} has shown that using intermediate checkpoints from the teacher's training process as an implicit ``curriculum'' for progressive distillation can significantly speed up training. However, such schemes require storing these checkpoints, and often require careful selection of the intermediate checkpoints to train on, which can be impractical for large-scale training. In this paper, we show that a curriculum can be \emph{extracted} from just the fully trained teacher network, and that this extracted curriculum can give similar efficiency benefits to those of progressive distillation. Our extraction scheme is natural; we use a random projection of the hidden representations of the teacher network to progressively train the student network, before training using the output of the full network. We show that our scheme significantly outperforms one-shot distillation and achieves a performance similar to that of progressive distillation for learning sparse parities with two-layer networks, and provide theoretical guarantees for this setting. Additionally, we show that our method outperforms one-shot distillation even when using transformer-based architectures, both for sparse-parity learning, and language modeling tasks.

Quantum kernels quantify similarity between data points by measuring the inner product between quantum states, computed through quantum circuit measurements. By embedding data into quantum systems, quantum kernel feature maps, that may be classically intractable to compute, could efficiently exploit high-dimensional Hilbert spaces to capture complex patterns. However, designing effective quantum feature maps remains a major challenge. Many quantum kernels, such as the fidelity kernel, suffer from exponential concentration, leading to near-identity kernel matrices that fail to capture meaningful data correlations and lead to overfitting and poor generalization. In this paper, we propose a novel strategy for constructing quantum kernels that achieve good generalization performance, drawing inspiration from benign overfitting in classical machine learning. Our approach introduces the concept of local-global quantum kernels, which combine two complementary components: a local quantum kernel based on measurements of small subsystems and a global quantum kernel derived from full-system measurements. Through numerical experiments, we demonstrate that local-global quantum kernels exhibit benign overfitting, supporting the effectiveness of our approach in enhancing quantum kernel methods.

The behavior of multivariate dynamical processes is often governed by underlying structural connections that relate the components of the system. For example, brain activity which is often measured via time series is determined by an underlying structural graph, where nodes represent neurons or brain regions and edges represent cortical connectivity. Existing methods for inferring structural connections from observed dynamics, such as correlation-based or spectral techniques, may fail to fully capture complex relationships in high-dimensional time series in an interpretable way. Here, we propose the use of path signatures--a mathematical framework that encodes geometric and temporal properties of continuous paths--to address this problem. Path signatures provide a reparametrization-invariant characterization of dynamical data and, in particular, can be used to compute the lead matrix which reveals lead-lag phenomena. We showcase our approach on time series from coupled oscillators in the Kuramoto model defined on a stochastic block model graph, termed the Kuramoto stochastic block model (KSBM). Using mean-field theory and Gaussian approximations, we analytically derive reduced models of KSBM dynamics in different temporal regimes and theoretically characterize the lead matrix in these settings. Leveraging these insights, we propose a novel signature-based community detection algorithm, achieving exact recovery of structural communities from observed time series in multiple KSBM instances. Our results demonstrate that path signatures provide a novel perspective on analyzing complex neural data and other high-dimensional systems, explicitly exploiting temporal functional relationships to infer underlying structure.

With the growing interest in quantum machine learning, the perceptron -- a fundamental building block in traditional machine learning -- has emerged as a valuable model for exploring quantum advantages. Two quantum perceptron algorithms based on Grover's search, were developed in arXiv:1602.04799 to accelerate training and improve statistical efficiency in perceptron learning. This paper points out and corrects a mistake in the proof of Theorem 2 in arXiv:1602.04799. Specifically, we show that the probability of sampling from a normal distribution for a $D$-dimensional hyperplane that perfectly classifies the data scales as $\Omega(\gamma^{D})$ instead of $\Theta({\gamma})$, where $\gamma$ is the margin. We then revisit two well-established linear programming algorithms -- the ellipsoid method and the cutting plane random walk algorithm -- in the context of perceptron learning, and show how quantum search algorithms can be leveraged to enhance the overall complexity. Specifically, both algorithms gain a sub-linear speed-up $O(\sqrt{N})$ in the number of data points $N$ as a result of Grover's algorithm and an additional $O(D^{1.5})$ speed-up is possible for cutting plane random walk algorithm employing quantum walk search.

Solving multiple parametrised related systems is an essential component of many numerical tasks. Borrowing strength from the solved systems and learning will make this process faster. In this work, we propose a novel probabilistic linear solver over the parameter space. This leverages information from the solved linear systems in a regression setting to provide an efficient posterior mean and covariance. We advocate using this as companion regression model for the preconditioned conjugate gradient method, and discuss the favourable properties of the posterior mean and covariance as the initial guess and preconditioner. We also provide several design choices for this companion solver. Numerical experiments showcase the benefits of using our novel solver in a hyperparameter optimisation problem.

Nearly all identifiability results in unsupervised representation learning inspired by, e.g., independent component analysis, factor analysis, and causal representation learning, rely on assumptions of additive independent noise or noiseless regimes. In contrast, we study the more general case where noise can take arbitrary forms, depend on latent variables, and be non-invertibly entangled within a nonlinear function. We propose a general framework for identifying latent variables in the nonparametric noisy settings. We first show that, under suitable conditions, the generative model is identifiable up to certain submanifold indeterminacies even in the presence of non-negligible noise. Furthermore, under the structural or distributional variability conditions, we prove that latent variables of the general nonlinear models are identifiable up to trivial indeterminacies. Based on the proposed theoretical framework, we have also developed corresponding estimation methods and validated them in various synthetic and real-world settings. Interestingly, our estimate of the true GDP growth from alternative measurements suggests more insightful information on the economies than official reports. We expect our framework to provide new insight into how both researchers and practitioners deal with latent variables in real-world scenarios.