Progressive Magnitude-Based Pruning Achieves Sparse Neural Networks in Single Training Cycle
Researchers propose a neural network pruning method that gradually increases sparsity during a single training cycle, rather than requiring multiple iterative cycles as in prior work. The approach uses magnitude-based pruning masks and demonstrates competitive or superior accuracy-sparsity tradeoffs compared to established methods like the Lottery Ticket Hypothesis on CIFAR-10 and MNIST benchmarks. This work is significant because it reduces computational cost while maintaining model performance, making neural network compression more practical for deployment.
A new study on arXiv presents progressive magnitude-based pruning as an efficient alternative to iterative neural network pruning methods. The technique gradually increases sparsity during training using a linear schedule and updates pruning masks based on active weight magnitudes, eliminating the need for multiple complete training cycles required by methods like the Lottery Ticket Hypothesis (LTH). Experiments on CIFAR-10 and MNIST across ResNet, VGG-style, and LeNet architectures show the method achieves 95.12% accuracy on ResNet-18 at 72.9% sparsity, outperforming LTH's 90.5%, and maintains accuracy within 0.1 percentage points of dense baselines across 70-85% sparsity levels. At extreme sparsity levels (97-98%), the method matches or exceeds performance of competing approaches like SNIP and GraSP. The results suggest progressive magnitude-based pruning offers a computationally efficient single-cycle approach for neural network sparsification.
What's missing
The study's evaluation is limited to relatively small-scale datasets (CIFAR-10, MNIST) and architectures; generalization to larger modern models (transformers, large language models) and ImageNet-scale datasets is not addressed. The paper does not discuss computational overhead of the pruning process itself or wall-clock training time comparisons with baseline methods.
What different sources said
- arXiv cs.LGCenter
Finding Sparse Subnetworks in One Training Cycle via Progressive Magnitude-Based Pruning
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.