Study Reveals Stability Challenges in Neural Network Growth During Training
Researchers found that growing new neurons during neural network training is fundamentally different from pruning, with newly added units receiving weaker gradient signals than established ones. The study shows that while growth can achieve high accuracy in simple tasks, pruning performs better in complex image-classification settings and when networks are retrained from scratch. The findings suggest that adaptive neural networks require careful timing and integration strategies for new units to be effective.
A new preprint from arXiv examines structural plasticity in deep learning—the ability to modify network architecture during training by adding or removing neurons. While pruning (removing units) and growth (adding units) might seem like inverse operations, the researchers demonstrate they behave quite differently. New neurons inserted into an already-optimized network suffer from a "backward-starvation" problem: they participate in forward computations but receive significantly weaker gradient signals during backpropagation compared to established units. This disadvantage is negligible in simple benchmarks but becomes pronounced in harder tasks like image classification with convolutional networks. The study shows that growth can achieve high final accuracy but underperforms pruning when performance is averaged across training or when sparse networks are retrained from scratch. Interventions targeting optimizer state and unit integration can help, but don't guarantee better final networks. The authors conclude that growth should be treated as a time-sensitive optimization process rather than a simple architecture-search operator.
What's missing
The study does not discuss computational costs or memory overhead of growth versus pruning during training. Additionally, the paper does not compare against other recent adaptive architecture methods beyond basic growth/pruning baselines, and the generalization of findings to other domains beyond image classification (e.g., NLP, reinforcement learning) remains unexplored.
What different sources said
- arXiv cs.LGCenter
On the Stability of Growth in Structural Plasticity
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.