Authors are listed in alphabetical order of last names. April 15, 2025 Abstract In this paper, we provide an overview of a common phenomenon, condensation, observed during the nonlinear training of neural networks: During the nonlinear training of neural networks, neurons in the same layer tend to condense into groups with similar outputs. Empirical observations suggest that the number of condensed clusters of neurons in the same layer typically increases monotonically as training progresses. Neural networks with small weight initializations or Dropout optimization can facilitate this condensation process. We also examine the underlying mechanisms of condensation from the perspectives of training dynamics and the structure of the loss landscape. The condensation phenomenon offers valuable insights into the generalization abilities of neural networks and correlates to stronger reasoning abilities in transformer-based language models. 1 Introduction Deep neural networks (DNNs) have demonstrated remarkable performance across a wide range of applications. In particular, scaling laws suggest that improvements in performance for Large Language Models (LLMs) are closely tied to the size of both the model and the dataset [KMH + 20]. Understanding how these large-scale neural networks achieve such extraordinary performance is crucial for developing principles that guide the design of more efficient, robust, and computationally cost-effective machine learning models. However, the study of large neural networks presents significant challenges, such as their enormous parameters and complex network architectures. Additionally, the data--ranging from language to image data--are often too complex to analyze using traditional methods.

Nesterov's accelerated gradient method (NAG) marks a pivotal advancement in gradient-based optimization, achieving faster convergence compared to the vanilla gradient descent method for convex functions. However, its algorithmic complexity when applied to strongly convex functions remains unknown, as noted in the comprehensive review by Chambolle and Pock [2016]. This issue, aside from the critical step size, was addressed by Li et al. [2024b], with the monotonic case further explored by Fu and Shi [2024]. In this paper, we introduce a family of controllable momentum coefficients for forward-backward accelerated methods, focusing on the critical step size $s=1/L$. Unlike traditional linear forms, the proposed momentum coefficients follow an $\alpha$-th power structure, where the parameter $r$ is adaptively tuned to $\alpha$. Using a Lyapunov function specifically designed for $\alpha$, we establish a controllable $O\left(1/k^{2\alpha} \right)$ convergence rate for the NAG-$\alpha$ method, provided that $r > 2\alpha$. At the critical step size, NAG-$\alpha$ achieves an inverse polynomial convergence rate of arbitrary degree by adjusting $r$ according to $\alpha > 0$. We further simplify the Lyapunov function by expressing it in terms of the iterative sequences $x_k$ and $y_k$, eliminating the need for phase-space representations. This simplification enables us to extend the controllable $O \left(1/k^{2\alpha} \right)$ rate to the monotonic variant, M-NAG-$\alpha$, thereby enhancing optimization efficiency. Finally, by leveraging the fundamental inequality for composite functions, we extended the controllable $O\left(1/k^{2\alpha} \right)$ rate to proximal algorithms, including the fast iterative shrinkage-thresholding algorithm (FISTA-$\alpha$) and its monotonic counterpart (M-FISTA-$\alpha$).

Bayesian optimization (BO) is an effective paradigm for the optimization of expensive-to-sample systems. Standard BO learns the performance of a system $f(x)$ by using a Gaussian Process (GP) model; this treats the system as a black-box and limits its ability to exploit available structural knowledge (e.g., physics and sparse interconnections in a complex system). Grey-box modeling, wherein the performance function is treated as a composition of known and unknown intermediate functions $f(x, y(x))$ (where $y(x)$ is a GP model) offers a solution to this limitation; however, generating an analytical probability density for $f$ from the Gaussian density of $y(x)$ is often an intractable problem (e.g., when $f$ is nonlinear). Previous work has handled this issue by using sampling techniques or by solving an auxiliary problem over an augmented space where the values of $y(x)$ are constrained by confidence intervals derived from the GP models; such solutions are computationally intensive. In this work, we provide a detailed implementation of a recently proposed grey-box BO paradigm, BOIS, that uses adaptive linearizations of $f$ to obtain analytical expressions for the statistical moments of the composite function. We show that the BOIS approach enables the exploitation of structural knowledge, such as that arising in interconnected systems as well as systems that embed multiple GP models and combinations of physics and GP models. We benchmark the effectiveness of BOIS against standard BO and existing grey-box BO algorithms using a pair of case studies focused on chemical process optimization and design. Our results indicate that BOIS performs as well as or better than existing grey-box methods, while also being less computationally intensive.

Bayesian optimization (BO) has proven to be an effective paradigm for the global optimization of expensive-to-sample systems. One of the main advantages of BO is its use of Gaussian processes (GPs) to characterize model uncertainty which can be leveraged to guide the learning and search process. However, BO typically treats systems as black-boxes and this limits the ability to exploit structural knowledge (e.g., physics and sparse interconnections). Composite functions of the form $f(x, y(x))$, wherein GP modeling is shifted from the performance function $f$ to an intermediate function $y$, offer an avenue for exploiting structural knowledge. However, the use of composite functions in a BO framework is complicated by the need to generate a probability density for $f$ from the Gaussian density of $y$ calculated by the GP (e.g., when $f$ is nonlinear it is not possible to obtain a closed-form expression). Previous work has handled this issue using sampling techniques; these are easy to implement and flexible but are computationally intensive. In this work, we introduce a new paradigm which allows for the efficient use of composite functions in BO; this uses adaptive linearizations of $f$ to obtain closed-form expressions for the statistical moments of the composite function. We show that this simple approach (which we call BOIS) enables the exploitation of structural knowledge, such as that arising in interconnected systems as well as systems that embed multiple GP models and combinations of physics and GP models. Using a chemical process optimization case study, we benchmark the effectiveness of BOIS against standard BO and sampling approaches. Our results indicate that BOIS achieves performance gains and accurately captures the statistics of composite functions.

Loss functions serve as the foundation of supervised learning and are often chosen prior to model development. To avoid potentially ad hoc choices of losses, statistical decision theory describes a desirable property for losses known as \emph{properness}, which asserts that Bayes' rule is optimal. Recent works have sought to \emph{learn losses} and models jointly. Existing methods do this by fitting an inverse canonical link function which monotonically maps $\mathbb{R}$ to $[0,1]$ to estimate probabilities for binary problems. In this paper, we extend monotonicity to maps between $\mathbb{R}^{C-1}$ and the projected probability simplex $\tilde{\Delta}^{C-1}$ by using monotonicity of gradients of convex functions. We present {\sc LegendreTron} as a novel and practical method that jointly learns \emph{proper canonical losses} and probabilities for multiclass problems. Tested on a benchmark of domains with up to 1,000 classes, our experimental results show that our method consistently outperforms the natural multiclass baseline under a $t$-test at 99% significance on all datasets with greater than 10 classes.

Bayesian Optimization (BO) is a technique for sample-efficient black-box optimization that employs probabilistic models to identify promising input locations for evaluation. When dealing with composite-structured functions, such as f=g o h, evaluating a specific location x yields observations of both the final outcome f(x) = g(h(x)) as well as the intermediate output(s) h(x). Previous research has shown that integrating information from these intermediate outputs can enhance BO performance substantially. However, existing methods struggle if the outputs h(x) are high-dimensional. Many relevant problems fall into this setting, including in the context of generative AI, molecular design, or robotics. To effectively tackle these challenges, we introduce Joint Composite Latent Space Bayesian Optimization (JoCo), a novel framework that jointly trains neural network encoders and probabilistic models to adaptively compress high-dimensional input and output spaces into manageable latent representations. This enables viable BO on these compressed representations, allowing JoCo to outperform other state-of-the-art methods in high-dimensional BO on a wide variety of simulated and real-world problems.

Bayesian Optimization (BO) is used to find the global optima of black box functions. In this work, we propose a practical BO method of function compositions where the form of the composition is known but the constituent functions are expensive to evaluate. By assuming an independent Gaussian process (GP) model for each of the constituent black-box function, we propose Expected Improvement (EI) and Upper Confidence Bound (UCB) based BO algorithms and demonstrate their ability to outperform not just vanilla BO but also the current state-of-art algorithms. We demonstrate a novel application of the proposed methods to dynamic pricing in revenue management when the underlying demand function is expensive to evaluate.

In this paper, we revisit the class of iterative shrinkage-thresholding algorithms (ISTA) for solving the linear inverse problem with sparse representation, which arises in signal and image processing. It is shown in the numerical experiment to deblur an image that the convergence behavior in the logarithmic-scale ordinate tends to be linear instead of logarithmic, approximating to be flat. Making meticulous observations, we find that the previous assumption for the smooth part to be convex weakens the least-square model. Specifically, assuming the smooth part to be strongly convex is more reasonable for the least-square model, even though the image matrix is probably ill-conditioned. Furthermore, we improve the pivotal inequality tighter for composite optimization with the smooth part to be strongly convex instead of general convex, which is first found in [Li et al., 2022]. Based on this pivotal inequality, we generalize the linear convergence to composite optimization in both the objective value and the squared proximal subgradient norm. Meanwhile, we set a simple ill-conditioned matrix which is easy to compute the singular values instead of the original blur matrix. The new numerical experiment shows the proximal generalization of Nesterov's accelerated gradient descent (NAG) for the strongly convex function has a faster linear convergence rate than ISTA. Based on the tighter pivotal inequality, we also generalize the faster linear convergence rate to composite optimization, in both the objective value and the squared proximal subgradient norm, by taking advantage of the well-constructed Lyapunov function with a slight modification and the phase-space representation based on the high-resolution differential equation framework from the implicit-velocity scheme.

Sometimes, you see a diagram and it gives you an'aha ha' moment I saw it on Frederick kratzert's blog Using the input variables x and y, The forwardpass (left half of the figure) calculates output z as a function of x and y i.e. f(x,y) The right side of the figures shows the backwardpass. Receiving dL/dz (the derivative of the total loss with respect to the output z), we can calculate the individual gradients of x and y on the loss function by applying the chain rule, as shown in the figure. This post is a part of my forthcoming book on Mathematical foundations of Data Science. The goal of the neural network is to minimise the loss function for the whole network of neurons. Hence, the problem of solving equations represented by the neural network also becomes a problem of minimising the loss function for the entire network.

The chain rule is an important derivative rule that allows us to work with composite functions. It is essential in understanding the workings of the backpropagation algorithm, which applies the chain rule extensively in order to calculate the error gradient of the loss function with respect to each weight of a neural network. We will be building on our earlier introduction to the chain rule, by tackling more challenging functions. In this tutorial, you will discover how to apply the chain rule of calculus to challenging functions. The Chain Rule of Calculus – Even More Functions Photo by Nan Ingraham, some rights reserved.