Reproduce “Generative NeuroEvolution for Deep Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Phillip Verbancsics, Josh Harguess
Researchers investigated the application of HyperNEAT for deep learning, demonstrating that its effectiveness in image classification is substantially enhanced when it acts as a feature extractor within a Convolutional Neural Network (CNN) architecture. This hybrid approach achieved 92.1% accuracy on the MNIST dataset when combined with backward propagation, outperforming direct classification by HyperNEAT alone.
Key insights
- Evolutionary algorithms, specifically generative neuro-evolution with indirect encodings, can effectively train complex weight patterns in deeper network architectures.
- A hybrid approach combining evolutionary feature learning via HyperNEAT with traditional supervised learning for classification significantly boosts performance over either method alone.
- The choice of underlying neural network architecture (e.g., CNN vs. traditional ANN) is critical for HyperNEAT's success, especially when exploiting geometric regularities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.