Study on rich regime training in deep learning, finding active parameters in bottom layers.
problem Understanding the practical success of deep learning models.
method Empirical study on rich regime training with benchmark datasets, re-initialization analysis, and probabilistic Layer-Wise Sparse SGD.
result Probabilistic Layer-Wise Sparse SGD matches vanilla SGD's generalization performance with improved efficiency.
To reduce the long training time of large deep neural network (DNN) models, distributed synchronous stochastic gradient descent (S-SGD) is commonly used on a cluster of workers. However, the speedup brought by multiple workers is limited by the communication overhead. Two approaches, namely pipelining and gradient spar…
Layer-wise preconditioning methods improve neural network optimization and feature learning.
problem Suboptimal feature learning in standard optimization algorithms.
method Layer-wise preconditioning methods that introduce preconditioners per axis of each layer's weight tensors.
result Layer-wise preconditioning is necessary for provable feature learning in linear and single-index models.
Stochastic gradient descent (SGD) has achieved great success in training deep neural network, where the gradient is computed through back-propagation. However, the back-propagated values of different layers vary dramatically. This inconsistence of gradient magnitude across different layers renders optimization of deep …
We propose NovoGrad, an adaptive stochastic gradient descent method with layer-wise gradient normalization and decoupled weight decay. In our experiments on neural networks for image classification, speech recognition, machine translation, and language modeling, it performs on par or better than well tuned SGD with mom…
A fundamental question in deep learning concerns the role played by individual layers in a deep neural network (DNN) and the transferable properties of the data representations which they learn. To the extent that layers have clear roles, one should be able to optimize them separately using layer-wise loss functions. S…
We propose a new optimization method for training feed-forward neural networks. By rewriting the activation function as an equivalent proximal operator, we approximate a feed-forward neural network by adding the proximal operators to the objective function as penalties, hence we call the lifted proximal operator machin…
Deep Neural Network (DNN) is powerful but computationally expensive and memory intensive, thus impeding its practical usage on resource-constrained front-end devices. DNN pruning is an approach for deep model compression, which aims at eliminating some parameters with tolerable performance degradation. In this paper, w…
Large SGD step sizes lead to sparse feature learning in neural networks.
problem Sparse feature learning in neural networks with large step sizes.
method Empirical observations and theoretical analysis of SGD dynamics.
result Large step sizes induce implicit regularization leading to sparse predictors.
SGD learns sparse parities near computational limits with discontinuous phase transitions.
problem Learning sparse parities in deep learning.
method Empirical and theoretical analysis of SGD on sparse parities.
result SGD makes progress on sparse parities via Fourier gap, not Langevin-like mechanism.
New analysis enables inversion of deep generative models with unique solutions.
problem Inverting deep generative models like GANs and VAEs.
method Sparse representation theory and layer-wise inversion pursuit algorithms.
result Invertible solutions for generative models with unique latent vectors.
In the analysis of machine learning models, it is often convenient to assume that the parameters are IID. This assumption is not satisfied when the parameters are updated through training processes such as SGD. A relaxation of the IID condition is a probabilistic symmetry known as exchangeability. We show the sense in …
Proposes a Bayesian approach for automatic node selection in sparse neural networks.
problem Reduces structural complexity and computational speedup in large-scale predictive models.
method Uses spike-and-slab Gaussian priors and variational Bayes approach for node selection.
result Establishes variational posterior consistency and optimal contraction rates for sparse networks.
New learning rate approach reveals phase transitions in SGD performance.
problem Understanding feature learning dynamics in neural networks.
method Characterizing the relationship between learning rate(s) and sample complexity for gradient-based algorithms.
result Phase transition from information exponent to generative exponent regime with different learning rates.
Dynamic Sparse Training finds efficient sparse networks from scratch.
problem Finding efficient sparse neural networks.
method Jointly optimizes network parameters and sparsity with trainable thresholds.
result Achieves state-of-the-art performance with minimal performance loss.
Paper solves k-sparse parity problem with sign SGD, matching SQ lower bound.
problem Solving k-sparse parity problems efficiently.
method Sign stochastic gradient descent on neural networks.
result Matches Statistical Query lower bound for solving k-sparse parity problems.
Noise in SGD affects overparameterized models, favoring sparse solutions.
problem Understanding and mitigating implicit bias in SGD with parameter-dependent noise.
method Theoretical analysis of a quadratically-parameterized model with label noise and Gaussian noise.
result SGD with label noise recovers sparse ground-truth solutions, while SGD with Gaussian noise overfits dense solutions.
Stochastic gradient descent (SGD) has been the dominant optimization method for training deep neural networks due to its many desirable properties. One of the more remarkable and least understood quality of SGD is that it generalizes relatively well on unseen data even when the neural network has millions of parameters…
A new pruning method finds sparse minimizers in flat regions of deep neural networks.
problem Finding sparse minimizers in flat regions of deep neural networks.
method Directional pruning using an ℓ1 proximal gradient algorithm. result Empirical results show promising sparsity (92%) and comparable minima to SGD.
Study on GD and SGD over diagonal networks, focusing on stepsizes and regularisation.
problem Understanding the impact of stochasticity and large stepsizes on gradient descent and SGD solutions.
method Investigation of GD and SGD over diagonal linear networks with macroscopic stepsizes, proving convergence and characterizing solutions.
result Large stepsizes consistently benefit SGD for sparse regression problems, but can hinder GD recovery of sparse solutions, especially in the edge of stability regime.
SGD performs optimally on dependent data, achieving both estimation and regret.
problem Performance of SGD on temporally dependent data.
method Stochastic gradient descent with martingale-type and sequential dependence.
result SGD achieves optimal estimation and regret under various dependence structures.
New property helps SGD learn sparse functions efficiently in neural networks.
problem Characterizing functions learnable by SGD in non-linear neural networks.
method Mean-field analysis, hierarchical merged-staircase property, dimension-free dynamics approximation.
result Merged-staircase property is necessary and nearly sufficient for SGD learnability.
SGD batch size affects autoencoder global minima sparsity and sharpness.
problem Investigating how batch size impacts autoencoder learning.
method Non-convex autoencoder training with SGD, varying batch sizes.
result SGD batch size influences global minimum sparsity and sharpness.
A new method reduces communication costs in distributed learning.
problem Reduces communication bottlenecks in distributed learning.
method Local SGD with communication-computation overlap and delay-corrected sparse model averaging.
result Theoretical convergence guarantees for smooth non-convex objectives.
Optimizes sparse fine-tuning for privacy in neural networks.
problem Performance gap between DP-SGD and non-private fine-tuning.
method Optimization-based approach using private gradient information for selecting trainable weights.
result Our selection method leads to better prediction accuracy compared to existing approaches.
Proposes QEP to mitigate quantization error propagation in layer-wise post-training quantization.
problem Growth of quantization errors across layers degrades performance, especially in low-bit regimes.
method Quantization Error Propagation (QEP) framework that explicitly propagates and compensates for quantization errors.
result QEP-enhanced layer-wise PTQ achieves substantially higher accuracy, especially in low-bit regimes.
Proposes spred for solving L1 penalty with SGD.
problem Solving L1 penalty in optimization problems. method Reparametrization and SGD approach.
result Proves spred as an exact differentiable solver of L1. Unified framework LPCD optimizes quantization of complex submodules.
problem Quantization of complex submodules in neural networks.
method Layer-Projected Coordinate Descent (LPCD) for quantizing arbitrary submodules.
result LPCD enhances both layer-wise PTQ methods and existing submodule approaches.
Layer-wise networks have a closed-form solution and a stopping criterion.
problem Training networks one layer at a time without backpropagation.
method Proved the Kernel Mean Embedding as the closed-form solution and developed a stopping criterion.
result Layer-wise networks converge to a highly desirable kernel for classification.
Layer-wise networks have a closed-form solution and a stopping criterion.
problem Training networks one layer at a time without backpropagation.
method Proved the closed-form solution using the kernel Mean Embedding and Neural Indicator Kernel.
result Layer-wise networks have a closed-form solution and a stopping criterion.
TempBalance improves neural network training by balancing learning rates across layers.
problem Improving neural network training performance.
method Layer-wise learning rate scheduling based on HT-SR Theory.
result TempBalance significantly outperforms ordinary SGD and state-of-the-art optimizers.
The paper provides statistical guarantees for SGD and ASGD in high-dimensional settings.
problem Theoretical understanding of SGD and ASGD in high-dimensional settings.
method Transfer of tools from high-dimensional time series to online learning, using coupling techniques.
result Established geometric-moment contraction and q-th moment convergence of SGD and ASGD. Paper analyzes D-SGD convergence with heterogeneous data and proposes topology learning.
problem Efficiently dealing with data heterogeneity in decentralized learning.
method Revisits D-SGD analysis, introduces neighborhood heterogeneity, and proposes topology learning.
result Formulates topology learning as a tractable optimization problem and demonstrates its effectiveness.
The stochastic gradient descent (SGD) algorithm has been widely used in statistical estimation for large-scale data due to its computational and memory efficiency. While most existing works focus on the convergence of the objective function or the error of the obtained solution, we investigate the problem of statistica…
Paper proposes new Bayesian neural network models for efficient learning.
problem Efficient learning and model compression in deep neural networks.
method Proposes Spike-and-Slab Group Lasso (SS-GL) and Spike-and-Slab Group Horseshoe (SS-GHS) priors for structured sparsity in Bayesian neural networks.
result Establishes competitive performance in prediction accuracy, model compression, and inference latency compared to baseline models.
In this paper, we tackle the problem of explanations in a deep-learning based model for recommendations by leveraging the technique of layer-wise relevance propagation. We use a Deep Convolutional Neural Network to extract relevant features from the input images before identifying similarity between the images in featu…
LEWIS merges LLMs without training, improving performance on specific tasks.
problem Limited performance improvement of merged models on specific benchmarks.
method Guided model merging using layer-wise sparsity and task-vector pruning.
result Improved model performance by up to 11.3% on math-solving tasks.
SGD tends to favor simpler subnetworks, improving generalization.
problem SGD's tendency to favor simpler subnetworks over complex ones.
method Identifying invariant sets and analyzing SGD's behavior around them.
result SGD collapses networks to simpler subnetworks, improving generalization.
Random layer-wise pruning profiles are as effective as metric-based ones for various datasets.
problem Reduction of model size and computational resources in neural networks.
method Conducted baseline experiments, developed RL-based search algorithm for finding transferable layer-wise pruning profiles.
result RL-based layer-wise pruning profiles are as good or better than best profiles found on the original dataset via exhaustive search.
DP-SGD can update fewer coordinates while maintaining privacy.
problem How to update fewer coordinates in DP-SGD without losing optimization signal.
method TP-TopK (Two-Phase TopK DP-SGD), a two-phase method for coordinate-sparse private training.
result Private training can update fewer coordinates without losing optimization signal, scaling noise with active dimension \(k\) instead of full dimension \(d\).
A long-standing problem in the theory of stochastic gradient descent (SGD) is to prove that its without-replacement version RandomShuffle converges faster than the usual with-replacement version. We present the first (to our knowledge) non-asymptotic solution to this problem, which shows that after a "reasonable" numbe…
EC2T creates sparse and ternary neural networks for resource-constrained devices.
problem Deploying deep neural networks on resource-constrained devices.
method Entropy-Constrained Trained Ternarization (EC2T) framework.
result EC2T creates sparse and ternary neural networks that are efficient in terms of storage and computation.
New algorithm optimizes linear system estimation from single trajectory.
problem Estimating LTI systems from a single trajectory.
method SGD with Reverse Experience Replay (SGD−RER) result Optimal guarantees for parameter and prediction errors.
3BASiL-TM decomposes LLMs into sparse and low-rank matrices for efficient compression.
problem Efficiently compressing large language models without significant performance loss.
method 3-Block ADMM method and transformer-matching refinement step for sparse plus low-rank decomposition.
result 3BASiL-TM reduces perplexity gap by over 30% and speeds up compression by 2.5x.
Analyzes layer-wise quantization effects in neural networks.
problem Identifying and fixing degradation in quantized neural networks.
method Layer-wise quantization analysis framework.
result Local fixes can significantly reduce quantization degradation.
Automatically learns flexible symmetry constraints in neural networks using gradients.
problem Fixed hard constraints on neural network functions that cannot be adapted.
method Improves parameterisations of soft equivariance and optimizes marginal likelihood using differentiable Laplace approximations.
result Achieves equivalent or improved performance on image classification tasks compared to baselines with hard-coded symmetry.
Stochastic gradient descent (SGD) is commonly used for optimization in large-scale machine learning problems. Langford et al. (2009) introduce a sparse online learning method to induce sparsity via truncated gradient. With high-dimensional sparse data, however, the method suffers from slow convergence and high variance…
Unified framework HASSLE-free decomposes large model weights into sparse and low-rank components.
problem Efficiently compress large foundation models to reduce inference costs.
method Designs a unified framework for sparse plus low-rank matrix decomposition with a local layer-wise reconstruction error objective.
result HASSLE-free framework significantly outperforms state-of-the-art methods in compression and evaluation benchmarks.