A new energy-efficient pruning method for federated learning.
problem Energy inefficiency in gradient sparsification for federated learning.
method Formalized energy-constrained projection problem and proposed Cost-Weighted Magnitude Pruning (CWMP).
result CWMP optimally balances performance and energy efficiency in federated learning.
Lookahead pruning extends single-layer optimization to multi-layer, outperforming magnitude-based pruning.
problem Pruning neural networks to reduce computational cost and memory usage.
method Developed a multi-layer optimization approach extending the single-layer optimization of magnitude-based pruning.
result Consistently outperforms magnitude-based pruning on various networks, especially in high sparsity.
New pruning method retains model expressiveness for NLP tasks.
problem Pruning large pretrained transformer models for real-world deployment.
method Mixture Gaussian Prior Pruning (MGPP) algorithm.
result MGPP outperforms existing pruning methods in high sparsity settings.
Pruning method removes less important features in linear models.
problem Removing less important features in linear models trained by gradient flow.
method Iterative Magnitude Pruning (IMP) applied to linear models trained by gradient flow.
result IMP prunes features with smallest projection onto the data.
Neural networks have achieved dramatic improvements in recent years and depict the state-of-the-art methods for many real-world tasks nowadays. One drawback is, however, that many of these models are overparameterized, which makes them both computationally and memory intensive. Furthermore, overparameterization can als…
Study examines effects of pruning techniques on deep learning models.
problem Understanding the impact of pruning methods on deep learning model structure and dynamics.
method Investigated differences in connectivity and learning dynamics of pruned models using various iterative pruning techniques.
result Emergence of structure in pruned models through magnitude-based unstructured pruning and weight rewinding.
Recent pruning methods at initialization fall short of random pruning's accuracy.
problem Improving neural network accuracy through pruning at initialization.
method Various pruning methods (SNIP, GraSP, SynFlow, magnitude pruning) are evaluated; per-layer pruning decisions are proposed.
result Randomly shuffling or sampling initial weights preserves or improves accuracy, suggesting challenges with pruning heuristics.
A new framework explains why early pruning works well.
problem Understanding why early pruning of neural networks leads to good performance.
method Gradient flow framework to unify pruning measures.
result Magnitude-based pruning removes least contributing parameters, leading to faster convergence.
This work explores the importance of model weights and Hessian bias in pruning.
problem Understanding the relative importance of model weights for efficient pruning.
method A principled exploration of pruning, focusing on linear models and neural networks.
result Asymptotic formulas reveal the performance of different pruning methods.
Pruned neural networks' error scales predictably with architecture and task.
problem Understanding the predictability of pruning across different scales and architectures.
method Functionally approximated the error of pruned networks, showing it is predictable in terms of invariant tying width, depth, and pruning level.
result The error of pruned networks follows a scaling law with interpretable coefficients that depend on architecture and task.
This work characterizes the fundamental limit of network pruning using statistical dimension and convex geometry.
problem The fundamental limit of network pruning is still lacking, especially for deep neural networks.
method Directly imposing sparsity constraint on the loss function and using statistical dimension in convex geometry.
result Characterizes the sharp phase transition point as the fundamental limit of pruning ratio.
New pruning methods improve dynamic sparse training performance.
problem Improving dynamic sparse training performance.
method Design and empirical analysis of pruning criteria.
result Most pruning methods yield similar results, but magnitude-based pruning performs best in low-density regimes.
The thesis explores IMP, a process that identifies winning tickets in DNNs, and its universality.
problem Understanding how winning subnetworks (tickets) perform across different problems.
method Iterative Magnitude Pruning (IMP) and comparison with Renormalisation Group (RG) theory.
result IMP identifies winning subnetworks that can perform similarly across various problems.
To address the limitations of existing magnitude-based pruning algorithms in cases where model weights or activations are of large and similar magnitude, we propose a novel perspective to discover parameter redundancy among channels and accelerate deep CNNs via channel pruning. Precisely, we argue that channels reveali…
Recent DNN pruning algorithms have succeeded in reducing the number of parameters in fully connected layers, often with little or no drop in classification accuracy. However, most of the existing pruning schemes either have to be applied during training or require a costly retraining procedure after pruning to regain c…
New pruning methods improve energy efficiency of neural networks.
problem Energy-efficient neural networks for devices with limited resources.
method Magnitude and Gradient based pruning at initialization and training of sparse architectures.
result Proposed novel pruning methods prevent full layer pruning and improve training.
IMP finds sparse subnetworks that match full networks, revealing geometric insights.
problem Finding sparse subnetworks that match full, overparameterized networks.
method Iterative magnitude pruning (IMP) algorithm, analyzing error landscape geometry.
result IMP masks found at end of training convey useful information for rewound networks.
Pruning is a standard technique for removing unnecessary structure from a neural network to reduce its storage footprint, computational demands, or energy consumption. Pruning can reduce the parameter-counts of many state-of-the-art neural networks by an order of magnitude without compromising accuracy, meaning these n…
The sophisticated structure of Convolutional Neural Network (CNN) allows for outstanding performance, but at the cost of intensive computation. As significant redundancies inevitably present in such a structure, many works have been proposed to prune the convolutional filters for computation cost reduction. Although ex…
The most common method for DNN pruning is hard thresholding of network weights, followed by retraining to recover any lost accuracy. Recently developed smart pruning algorithms use the DNN response over the training set for a variety of cost functions to determine redundant network weights, leading to less accuracy deg…
Sparser Random Feature Models via IMP (ShRIMP) efficiently learns sparse models for high-dimensional data.
problem Learning sparse models for high-dimensional data with sparse variable dependencies.
method Iterative Magnitude Pruning applied to Random Feature Models.
result ShRIMP achieves better or competitive test accuracy compared to state-of-the-art methods.
i-SpaSP prunes neural networks by identifying important groups of parameters, improving pruning efficiency.
problem Pruning neural networks to reduce computational cost and improve performance.
method i-SpaSP uses sparse signal recovery principles to iteratively identify and threshold important parameter groups.
result i-SpaSP achieves strong empirical results and theoretical convergence guarantees, improving pruning efficiency.
Our work connects parameter magnitudes and Hessian eigenspaces in deep neural nets.
problem Understanding the relationship between parameter magnitudes and Hessian curvature in deep learning models.
method Developed a matrix-free algorithm based on sketched SVDs to measure similarity between parameter masks and Hessian eigenspaces.
result Top Hessian eigenvectors tend to be concentrated around larger parameters, indicating a connection between parameter magnitudes and loss curvature.
Optimizes pruning masks for neural networks using probabilistic fine-tuning and PAC-Bayes bounds.
problem Improving neural network performance through adaptive pruning of weights.
method Optimizes stochastic pruning masks by minimizing expected loss, considering data-adaptive regularization and feature alignment.
result Probabilistic fine-tuning leads to improved test error over baseline methods in neural networks.
A new method prunes neural networks faster and more efficiently.
problem Reducing training time and memory usage for neural networks.
method Set-based Task-Adaptive Meta Pruning (STAMP) that meta-learns a pruning mask.
result Significantly improved compression rates and faster training speed.
Pruning FCNs reveals sub-networks that match CNNs' performance.
problem Understanding the inductive bias of pruning in neural networks.
method Iterative magnitude pruning of a simple FCN followed by analysis of the resulting architecture.
result Pruned FCNs exhibit key features of CNNs, suggesting new architectural biases.
RicciNets prunes neural networks by removing edges of low importance based on Ricci curvature, reducing FLOPs by 35%.
problem Pruning neural networks to reduce computational load and improve efficiency.
method RicciNets uses Ricci curvature to prune edges of low importance in a randomly wired neural network, reducing FLOPs.
result Reduction of almost 35% in FLOPs with no performance degradation.
Neural networks are known to be vulnerable to adversarial examples. Carefully chosen perturbations to real images, while imperceptible to humans, induce misclassification and threaten the reliability of deep learning systems in the wild. To guard against adversarial examples, we take inspiration from game theory and ca…
Sparse Transformers degrade semantic information first, with early layers encoding more.
problem Understanding how sparse Transformers affect learned representations and semantic information.
method Probed Transformers with progressively pruned weights to observe changes in semantic information and model behavior.
result Complex semantic information is first to degrade in sparse Transformers, with early layers encoding more.
HALO learns to prune neural networks by adaptively shrinking weights.
problem Sparsity and model size in deep neural networks.
method Bayesian hierarchical models and trainable parameters for adaptive sparsification.
result HALO learns to create highly sparse networks with significant performance gains.
New Lipschitz bound for ReLU networks resists weight rescaling.
problem Lack of robustness guarantees for ReLU networks under weight perturbations.
method Rescaling-invariant Lipschitz bound based on path-metrics.
result The new bound applies to various ReLU-DAG architectures and resists neuron-wise rescalings.
Paper proposes a method to prune neural networks, reducing storage and computation costs.
problem Reduction of storage and computational costs for deep neural networks.
method Statistical analysis of component significance using F-statistic-based screening technique.
result Pruned models are highly competitive with state-of-the-art approaches.
New method prunes large causal bounds LPs for scalable inference.
problem Computing causal bounds on graphs with unobserved confounders.
method Pruning LP formulations for scalability, extending to fractional LPs.
result Significant runtime improvement and scalable inference for large problems.
TENP prunes experts and neurons in Mixture-of-Experts models for efficient deployment.
problem Efficient deployment of large language models constrained by static parameter footprint.
method Structured Trapezoidal ExpertNeuron Pruning (TENP) identifies and retains important experts and neurons.
result DeepSeek model achieves 10% better performance on code generation tasks with 40% expert sparsity.
Pruning is a well-established technique for removing unnecessary structure from neural networks after training to improve the performance of inference. Several recent results have explored the possibility of pruning at initialization time to provide similar benefits during training. In particular, the "lottery ticket h…
Sparsity helps reduce the computational complexity of deep neural networks by skipping zeros. Taking advantage of sparsity is listed as a high priority in next generation DNN accelerators such as TPU. The structure of sparsity, i.e., the granularity of pruning, affects the efficiency of hardware accelerator design as w…
There has recently been an increasing desire to evaluate neural networks locally on computationally-limited devices in order to exploit their recent effectiveness for several applications; such effectiveness has nevertheless come together with a considerable increase in the size of modern neural networks, which constit…
CoDeQ simplifies joint model compression by integrating pruning and quantization.
problem Joint pruning and quantization methods are complex and require additional procedures.
method CoDeQ uses a dead-zone quantizer to directly induce sparsity and learn quantization parameters.
result CoDeQ achieves high sparsity and low-precision accuracy with minimal bit operations.
PoET-BiN reduces power consumption in neural networks on embedded devices.
problem Power inefficiency in neural network implementations on embedded platforms.
method Look-Up Table based implementation with a modified Decision Tree approach.
result Near state-of-the-art results with up to 6 orders of magnitude energy reduction.
New algorithm discovers causal graphs efficiently from observational data.
problem Discovering causal graphs from observational data efficiently.
method Approximating the score function using machine learning and applying scalable techniques.
result DAS algorithm reduces complexity and achieves competitive accuracy.
The recently proposed Lottery Ticket Hypothesis of Frankle and Carbin (2019) suggests that the performance of over-parameterized deep networks is due to the random initialization seeding the network with a small fraction of favorable weights. These weights retain their dominant status throughout training -- in a very r…
Lottery tickets find good initializations for IMP with sparse training.
problem Finding good initializations for iterative magnitude pruning (IMP) in sparse networks.
method Empirical study of IMP performance with varying pre-training data and iterations.
result Training on a small fraction of data suffices to obtain good initializations for IMP.
We rigorously evaluate three state-of-the-art techniques for inducing sparsity in deep neural networks on two large-scale learning tasks: Transformer trained on WMT 2014 English-to-German, and ResNet-50 trained on ImageNet. Across thousands of experiments, we demonstrate that complex techniques (Molchanov et al., 2017;…
Powerpropagation makes neural networks inherently sparse.
problem Training sparse neural networks to reduce computational footprint and model size.
method Introduces a new weight-parameterisation technique exploiting gradient descent dynamics.
result Models trained with Powerpropagation have a higher density of zero weights, allowing for more efficient pruning.
Long short-term memory (LSTM) has been widely used for sequential data modeling. Researchers have increased LSTM depth by stacking LSTM cells to improve performance. This incurs model redundancy, increases run-time delay, and makes the LSTMs more prone to overfitting. To address these problems, we propose a hidden-laye…
Real-world machine learning applications often have complex test metrics, and may have training and test data that are not identically distributed. Motivated by known connections between complex test metrics and cost-weighted learning, we propose addressing these issues by using a weighted loss function with a standard…
A machine learning configuration refers to a combination of preprocessor, learner, and hyperparameters. Given a set of configurations and a large dataset randomly split into training and testing set, we study how to efficiently select the best configuration with approximately the highest testing accuracy when trained f…
ICE-Pruning accelerates deep neural network pruning by 9.61x.
problem Efficiently pruning deep neural networks while maintaining accuracy.
method Iterative pruning with automatic fine-tuning steps, freezing strategy, and custom learning rate scheduler.
result Significantly reduces pruning time by up to 9.61x.