Despite the empirical success of Low-Rank Adaptation (LoRA) in fine-tuning pre-trained models, there is little theoretical understanding of how first-order methods with carefully crafted initialization adapt models to new tasks. In this work, we take the first step towards bridging this gap by theoretically analyzing the learning dynamics of LoRA for matrix factorization (MF) under gradient flow (GF), emphasizing the crucial role of initialization. For small initialization, we theoretically show that GF converges to a neighborhood of the optimal solution, with smaller initialization leading to lower final error. Our analysis shows that the final error is affected by the misalignment between the singular spaces of the pre-trained model and the target matrix, and reducing the initialization scale improves alignment. To address this misalignment, we propose a spectral initialization for LoRA in MF and theoretically prove that GF with small spectral initialization converges to the fine-tuning task with arbitrary precision. Numerical experiments from MF and image classification validate our findings.

Diffusion models are widely used for image editing tasks. Existing editing methods often design a representation manipulation procedure by curating an edit direction in the text embedding or score space. However, such a procedure faces a key challenge: overestimating the edit strength harms visual consistency while underestimating it fails the editing task. Notably, each source image may require a different editing strength, and it is costly to search for an appropriate strength via trial-and-error. To address this challenge, we propose ${\boldsymbol{Co}}{\text{ncept}}\,{\boldsymbol{Lan}}{\text{cet}}$ (CoLan), a zero-shot plug-and-play framework for principled representation manipulation in diffusion-based image editing. At inference time, we decompose the source input in the latent (text embedding or diffusion score) space as a sparse linear combination of the representations of the collected visual concepts. This allows us to accurately estimate the presence of concepts in each image, which informs the edit. Based on the editing task (replace/add/remove), we perform a customized concept transplant process to impose the corresponding editing direction. To sufficiently model the concept space, we curate a conceptual representation dataset, CoLan-150K, which contains diverse descriptions and scenarios of visual terms and phrases for the latent dictionary. Experiments on multiple diffusion-based image editing baselines show that methods equipped with CoLan achieve state-of-the-art performance in editing effectiveness and consistency preservation.

The problem of generating a perpetual dynamic scene from a single view is an important problem with widespread applications in augmented and virtual reality, and robotics. However, since dynamic scenes regularly change over time, a key challenge is to ensure that different generated views be consistent with the underlying 3D motions. Prior work learns such consistency by training on multiple views, but the generated scene regions often interpolate between training views and fail to generate perpetual views. To address this issue, we propose DynamicVoyager, which reformulates dynamic scene generation as a scene outpainting problem with new dynamic content. As 2D outpainting models struggle at generating 3D consistent motions from a single 2D view, we enrich 2D pixels with information from their 3D rays that facilitates learning of 3D motion consistency. More specifically, we first map the single-view video input to a dynamic point cloud using the estimated video depths. We then render a partial video of the point cloud from a novel view and outpaint the missing regions using ray information (e.g., the distance from a ray to the point cloud) to generate 3D consistent motions. Next, we use the outpainted video to update the point cloud, which is used for outpainting the scene from future novel views. Moreover, we can control the generated content with the input text prompt. Experiments show that our model can generate perpetual scenes with consistent motions along fly-through cameras. Project page: https://tianfr.github.io/DynamicVoyager.

Deep learning-based classifiers are known to be vulnerable to adversarial attacks. Existing methods for defending against such attacks require adding a defense mechanism or modifying the learning procedure (e.g., by adding adversarial examples). This paper shows that for certain data distributions one can learn a provably robust classifier using standard learning methods and without adding a defense mechanism. More specifically, this paper addresses the problem of finding a robust classifier for a binary classification problem in which the data comes from an isotropic mixture of Gaussians with orthonormal cluster centers. First, we characterize the largest $\ell_2$-attack any classifier can defend against while maintaining high accuracy, and show the existence of optimal robust classifiers achieving this maximum $\ell_2$-robustness. Next, we show that given data from the orthonormal Gaussian mixture model, gradient flow on a two-layer network with a polynomial ReLU activation and without adversarial examples provably finds an optimal robust classifier.

Classical optimisation theory guarantees monotonic objective decrease for gradient descent (GD) when employed in a small step size, or "stable", regime. In contrast, gradient descent on neural networks is frequently performed in a large step size regime called the "edge of stability", in which the objective decreases non-monotonically with an observed implicit bias towards flat minima. In this paper, we take a step toward quantifying this phenomenon by providing convergence rates for gradient descent with large learning rates in an overparametrised least squares setting. The key insight behind our analysis is that, as a consequence of overparametrisation, the set of global minimisers forms a Riemannian manifold $M$, which enables the decomposition of the GD dynamics into components parallel and orthogonal to $M$. The parallel component corresponds to Riemannian gradient descent on the objective sharpness, while the orthogonal component corresponds to a quadratic dynamical system. This insight allows us to derive convergence rates in three regimes characterised by the learning rate size: the subcritical regime, in which transient instability is overcome in finite time before linear convergence to a suboptimally flat global minimum; the critical regime, in which instability persists for all time with a power-law convergence toward the optimally flat global minimum; the supercritical regime, in which instability persists for all time with linear convergence to an oscillation of period two centred on the optimally flat global minimum.

Among many mysteries behind the success of deep networks lies the exceptional discriminative power of their learned representations as manifested by the intriguing Neural Collapse (NC) phenomenon, where simple feature structures emerge at the last layer of a trained neural network. Prior works on the theoretical understandings of NC have focused on analyzing the optimization landscape of matrix-factorization-like problems by considering the last-layer features as unconstrained free optimization variables and showing that their global minima exhibit NC. In this paper, we show that gradient flow on a two-layer ReLU network for classifying orthogonally separable data provably exhibits NC, thereby advancing prior results in two ways: First, we relax the assumption of unconstrained features, showing the effect of data structure and nonlinear activations on NC characterizations. Second, we reveal the role of the implicit bias of the training dynamics in facilitating the emergence of NC.

This paper studies the problem of training a two-layer ReLU network for binary classification using gradient flow with small initialization. We consider a training dataset with well-separated input vectors: Any pair of input data with the same label are positively correlated, and any pair with different labels are negatively correlated. Our analysis shows that, during the early phase of training, neurons in the first layer try to align with either the positive data or the negative data, depending on its corresponding weight on the second layer. A careful analysis of the neurons' directional dynamics allows us to provide an $\mathcal{O}(\frac{\log n}{\sqrt{\mu}})$ upper bound on the time it takes for all neurons to achieve good alignment with the input data, where $n$ is the number of data points and $\mu$ measures how well the data are separated. After the early alignment phase, the loss converges to zero at a $\mathcal{O}(\frac{1}{t})$ rate, and the weight matrix on the first layer is approximately low-rank. Numerical experiments on the MNIST dataset illustrate our theoretical findings.

The implicit bias of gradient-based training algorithms has been considered mostly beneficial as it leads to trained networks that often generalize well. However, Frei et al. (2023) show that such implicit bias can harm adversarial robustness. Specifically, they show that if the data consists of clusters with small inter-cluster correlation, a shallow (two-layer) ReLU network trained by gradient flow generalizes well, but it is not robust to adversarial attacks of small radius. Moreover, this phenomenon occurs despite the existence of a much more robust classifier that can be explicitly constructed from a shallow network. In this paper, we extend recent analyses of neuron alignment to show that a shallow network with a polynomial ReLU activation (pReLU) trained by gradient flow not only generalizes well but is also robust to adversarial attacks. Our results highlight the importance of the interplay between data structure and architecture design in the implicit bias and robustness of trained networks.

Recent theoretical analyses of the convergence of gradient descent (GD) to a global minimum for over-parametrized neural networks make strong assumptions on the step size (infinitesimal), the hidden-layer width (infinite), or the initialization (spectral, balanced). In this work, we relax these assumptions and derive a linear convergence rate for two-layer linear networks trained using GD on the squared loss in the case of finite step size, finite width and general initialization. Despite the generality of our analysis, our rate estimates are significantly tighter than those of prior work. Moreover, we provide a time-varying step size rule that monotonically improves the convergence rate as the loss function decreases to zero. Numerical experiments validate our findings.

In this paper, we analyze the convergence of gradient flow on a multi-layer linear model with a loss function of the form $f(W_1W_2\cdots W_L)$. We show that when $f$ satisfies the gradient dominance property, proper weight initialization leads to exponential convergence of the gradient flow to a global minimum of the loss. Moreover, the convergence rate depends on two trajectory-specific quantities that are controlled by the weight initialization: the *imbalance matrices*, which measure the difference between the weights of adjacent layers, and the least singular value of the *weight product* $W=W_1W_2\cdots W_L$. Our analysis exploits the fact that the gradient of the overparameterized loss can be written as the composition of the non-overparametrized gradient with a time-varying (weight-dependent) linear operator whose smallest eigenvalue controls the convergence rate. The key challenge we address is to derive a uniform lower bound for this time-varying eigenvalue that lead to improved rates for several multi-layer network models studied in the literature.

Neural networks trained via gradient descent with random initialization and without any regularization enjoy good generalization performance in practice despite being highly overparametrized. A promising direction to explain this phenomenon is to study how initialization and overparametrization affect convergence and implicit bias of training algorithms. In this paper, we present a novel analysis of single-hidden-layer linear networks trained under gradient flow, which connects initialization, optimization, and overparametrization. Firstly, we show that the squared loss converges exponentially to its optimum at a rate that depends on the level of imbalance of the initialization. Secondly, we show that proper initialization constrains the dynamics of the network parameters to lie within an invariant set. In turn, minimizing the loss over this set leads to the min-norm solution. Finally, we show that large hidden layer width, together with (properly scaled) random initialization, ensures proximity to such an invariant set during training, allowing us to derive a novel non-asymptotic upper-bound on the distance between the trained network and the min-norm solution.

A challenge of pan/tilt/zoom (PTZ) camera networks for efficient and flexible visual monitoring is automated active network reconfiguration in response to environmental stimuli. In this paper, given an event/activity distribution over a convex environment, we propose a new provably correct reactive coverage control algorithm for PTZ camera networks that continuously (re) configures camera orientations and zoom levels (i.e., angles of view) in order to locally maximize their total coverage quality. Our construction is based on careful modeling of visual sensing quality that is consistent with the physical nature of cameras, and we introduce a new notion of conic Voronoi diagrams, based on our sensing quality measures, to solve the camera network allocation problem: that is, to determine where each camera should focus in its field of view given all the other cameras' configurations. Accordingly, we design simple greedy gradient algorithms for both continuous-and discrete-time first-order PTZ camera dynamics that asymptotically converge a locally optimal coverage configuration. Finally, we provide numerical and experimental evidence demonstrating the effectiveness of the proposed coverage algorithms.