Gradient dynamics for low-rank fine-tuning beyond kernels

• URL: http://arxiv.org/abs/2411.15385v1
• Date: Sat, 23 Nov 2024 00:00:28 GMT
• Title: Gradient dynamics for low-rank fine-tuning beyond kernels
• Authors: Arif Kerem Dayi, Sitan Chen,
• Abstract summary: We study low-rank fine-tuning in a student-teacher setting. We prove under mild assumptions that a student model which is matrix at the base model and trained with online gradient descent will converge to the teacher.
• Score: 9.275532709125242
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: LoRA has emerged as one of the de facto methods for fine-tuning foundation models with low computational cost and memory footprint. The idea is to only train a low-rank perturbation to the weights of a pre-trained model, given supervised data for a downstream task. Despite its empirical sucess, from a mathematical perspective it remains poorly understood what learning mechanisms ensure that gradient descent converges to useful low-rank perturbations. In this work we study low-rank fine-tuning in a student-teacher setting. We are given the weights of a two-layer base model $f$, as well as i.i.d. samples $(x,f^*(x))$ where $x$ is Gaussian and $f^*$ is the teacher model given by perturbing the weights of $f$ by a rank-1 matrix. This generalizes the setting of generalized linear model (GLM) regression where the weights of $f$ are zero. When the rank-1 perturbation is comparable in norm to the weight matrix of $f$, the training dynamics are nonlinear. Nevertheless, in this regime we prove under mild assumptions that a student model which is initialized at the base model and trained with online gradient descent will converge to the teacher in $dk^{O(1)}$ iterations, where $k$ is the number of neurons in $f$. Importantly, unlike in the GLM setting, the complexity does not depend on fine-grained properties of the activation's Hermite expansion. We also prove that in our setting, learning the teacher model "from scratch'' can require significantly more iterations.

Related papers

• Conditional regression for the Nonlinear Single-Variable Model [4.565636963872865]
  We consider a model $F(X):=f(Pi_gamma):mathbbRdto[0,rmlen_gamma]$ where $Pi_gamma: [0,rmlen_gamma]tomathbbRd$ and $f:[0,rmlen_gamma]tomathbbR1$. We propose a nonparametric estimator, based on conditional regression, and show that it can achieve the $one$-dimensional optimal min-max rate
  arXiv  Detail & Related papers  (2024-11-14T18:53:51Z)
• Learning Orthogonal Multi-Index Models: A Fine-Grained Information Exponent Analysis [45.05072391903122]
  Information exponent plays important role in predicting sample complexity of online gradient descent. For multi-index models, focusing solely on the lowest degree can miss key structural details. We show that by considering both second- and higher-order terms, we can first learn the relevant space via the second-order terms.
  arXiv  Detail & Related papers  (2024-10-13T00:14:08Z)
• Scaling Laws in Linear Regression: Compute, Parameters, and Data [86.48154162485712]
  We study the theory of scaling laws in an infinite dimensional linear regression setup. We show that the reducible part of the test error is $Theta(-(a-1) + N-(a-1)/a)$. Our theory is consistent with the empirical neural scaling laws and verified by numerical simulation.
  arXiv  Detail & Related papers  (2024-06-12T17:53:29Z)
• A Unified Framework for Uniform Signal Recovery in Nonlinear Generative Compressed Sensing [68.80803866919123]
  Under nonlinear measurements, most prior results are non-uniform, i.e., they hold with high probability for a fixed $mathbfx*$ rather than for all $mathbfx*$ simultaneously. Our framework accommodates GCS with 1-bit/uniformly quantized observations and single index models as canonical examples. We also develop a concentration inequality that produces tighter bounds for product processes whose index sets have low metric entropy.
  arXiv  Detail & Related papers  (2023-09-25T17:54:19Z)
• A Unified Scheme of ResNet and Softmax [8.556540804058203]
  We provide a theoretical analysis of the regression problem: $| langle exp(Ax) + A x, bf 1_n rangle-1 ( exp(Ax) + Ax ) This regression problem is a unified scheme that combines softmax regression and ResNet, which has never been done before.
  arXiv  Detail & Related papers  (2023-09-23T21:41:01Z)
• Spatially heterogeneous learning by a deep student machine [0.0]
  Deep neural networks (DNN) with a huge number of adjustable parameters remain largely black boxes. We study supervised learning by a DNN of width $N$ and depth $L$ consisting of $NL$ perceptrons with $c$ inputs by a statistical mechanics approach called the teacher-student setting. We show that the problem becomes exactly solvable in what we call as 'dense limit': $N gg c gg 1$ and $M gg 1$ with fixed $alpha=M/c$ using the replica method developed in (H. Yoshino, (
  arXiv  Detail & Related papers  (2023-02-15T01:09:03Z)
• High-dimensional Asymptotics of Feature Learning: How One Gradient Step Improves the Representation [89.21686761957383]
  We study the first gradient descent step on the first-layer parameters $boldsymbolW$ in a two-layer network. Our results demonstrate that even one step can lead to a considerable advantage over random features.
  arXiv  Detail & Related papers  (2022-05-03T12:09:59Z)
• Locality defeats the curse of dimensionality in convolutional teacher-student scenarios [69.2027612631023]
  We show that locality is key in determining the learning curve exponent $beta$. We conclude by proving, using a natural assumption, that performing kernel regression with a ridge that decreases with the size of the training set leads to similar learning curve exponents to those we obtain in the ridgeless case.
  arXiv  Detail & Related papers  (2021-06-16T08:27:31Z)
• Model-Based Reinforcement Learning with Value-Targeted Regression [48.92439657407732]
  We focus on finite-horizon episodic RL where the transition model $P$ belongs to a known family of models $mathcalP$. We derive a bound on the regret, which, in the special case of linear mixtures, the regret bound takes the form $tildemathcalO(dsqrtH3T)$.
  arXiv  Detail & Related papers  (2020-06-01T17:47:53Z)
• Learning nonlinear dynamical systems from a single trajectory [102.60042167341956]
  We introduce algorithms for learning nonlinear dynamical systems of the form $x_t+1=sigma(Thetastarx_t)+varepsilon_t$. We give an algorithm that recovers the weight matrix $Thetastar$ from a single trajectory with optimal sample complexity and linear running time.
  arXiv  Detail & Related papers  (2020-04-30T10:42:48Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.