In this paper, we revisit and improve the convergence of policy gradient (PG), natural PG (NPG) methods, and their variance-reduced variants, under general smooth policy parametrizations.

Being infinite dimensional, non-parametric information geometry has long faced an "intractability barrier" due to the fact that the Fisher-Rao metric is now a functional incurring difficulties in defining its inverse. This paper introduces a novel framework to resolve the intractability with an Orthogonal Decomposition of the Tangent Space ($T_fM = S \oplus S^{\perp}$), where $S$ represents an observable covariate subspace. Through the decomposition, we derive the Covariate Fisher Information Matrix (cFIM), denoted as ${\bf G}_f$, which is a finite-dimensional and computable representative of information extractable from the manifold's geometry. Significantly, by proving the Trace Theorem: $H_G(f) = \text{Tr}({\bf G}_f)$, we establish a rigorous foundation for the G-entropy previously introduced by us, thereby identifying it as a fundamental geometric invariant representing the total explainable statistical information captured by the probability distribution associated with a model. Furthermore, we establish a link between ${\bf G}_f$ and the second derivative (i.e. the curvature) of the KL-divergence, leading to the notion of Covariate Cramér-Rao Lower Bound(CRLB). We demonstrate that ${\bf G}_f$ is congruent to the Efficient Fisher Information Matrix, thereby providing fundamental limits of variance for semi-parametric estimators. Finally, we apply our geometric framework to the Manifold Hypothesis, lifting the latter from a heuristic assumption into a testable condition of rank-deficiency within the cFIM. By defining the Information Capture Ratio, we provide a rigorous method for estimating intrinsic dimensionality in high-dimensional data. In short, our work bridges the gap between abstract information geometry and the demand of explainable AI, by providing a tractable path for assessing the statistical coverage and the efficiency of non-parametric models.

Approximate Natural Gradient Descent (NGD) methods are an important family of optimisers for deep learning models, which use approximate Fisher information matrices to pre-condition gradients during training. The empirical Fisher (EF) method approximates the Fisher information matrix empirically by reusing the per-sample gradients collected during back-propagation. Despite its ease of implementation, the EF approximation has its theoretical and practical limitations. This paper investigates the issue of EF, which is shown to be a major cause of its poor empirical approximation quality. An improved empirical Fisher (iEF) method is proposed to address this issue, which is motivated as a generalised NGD method from a loss reduction perspective, meanwhile retaining the practical convenience of EF.

We provide quantitative bounds measuring the $L^2$ difference in function space between the trajectory of a finite-width network trained on finitely many samples from the idealized kernel dynamics of infinite width and infinite data. An implication of the bounds is that the network is biased to learn the top eigenfunctions of the Neural Tangent Kernel not just on the training set but over the entire input space. This bias depends on the model architecture and input distribution alone and thus does not depend on the target function which does not need to be in the RKHS of the kernel. The result is valid for deep architectures with fully connected, convolutional, and residual layers. Furthermore the width does not need to grow polynomially with the number of samples in order to obtain high probability bounds up to a stopping time. The proof exploits the low-effective-rank property of the Fisher Information Matrix at initialization, which implies a low effective dimension of the model (far smaller than the number of parameters). We conclude that local capacity control from the low effective rank of the Fisher Information Matrix is still underexplored theoretically.

Second-order optimization algorithms, such as the Newton method and the natural gradient descent (NGD) method exhibit excellent convergence properties for training deep neural networks, but the high computational cost limits its practical application. In this paper, we focus on the NGD method and propose a novel layer-wise natural gradient descent (LNGD) method to further reduce computational costs and accelerate the training process. Specifically, based on the block diagonal approximation of the Fisher information matrix, we first propose the layer-wise sample method to compute each block matrix without performing a complete back-propagation. Then, each block matrix is approximated as a Kronecker product of two smaller matrices, one of which is a diagonal matrix, while keeping the traces equal before and after approximation. By these two steps, we provide a new approximation for the Fisher information matrix, which can effectively reduce the computational cost while preserving the main information of each block matrix. Moreover, we propose a new adaptive layer-wise learning rate to further accelerate training. Based on these new approaches, we propose the LNGD optimizer. The global convergence analysis of LNGD is established under some assumptions. Experiments on image classification and machine translation tasks show that our method is quite competitive compared to the state-of-the-art methods.

Recent advances using Kernel Logistic Regression (KLR) have demonstrated that learning can sculpt these landscapes to achieve capacities far exceeding classical limits [1-3]. Our previous phenomenological analysis identified a Ridge of Optimization where stability is maximized via a mechanism we termed Spectral Concentration, defined as a state where the weight spectrum exhibits a sharp hierarchy [4]. However, a deeper question remains: Why does the learning dynamics self-organize into this specific spectral state? Why does the system operate at the brink of instability? T o answer these questions, we must look beyond the Euclidean geometry of the weight parameters and consider the intrinsic geometry of the probability distributions they represent. This is the domain of Information Geometry [5]. In this work, we reinterpret the KLR Hopfield network as a statistical manifold equipped with a Fisher-Rao metric.