Reference library
Safe Crossover of Neural Networks Through Neuron Alignment
Paper summary
The authors align hidden neurons using their activation responses to common data, with pairwise correlation or CCA-derived matching, then interpolate consistently reordered weights. Tests on 200 independently trained network pairs per dataset show less damage than unaligned interpolation. On CIFAR-10, midpoint accuracy is approximately 25% with alignment versus 10% naïve, below parents near 40%. These are fixed-architecture feedforward interpolation experiments, rather than whole evolutionary search comparisons.
Relevance to the project
Assistant interpretation: supports matching inherited neural components before coefficient mixing and consistently transforming incident weights/biases. It does not validate splicing recurrent d representations or online-plastic type controllers.
Proof of concept using Adam-trained fully connected same-topology models on MNIST/CIFAR-10. Semi-matching can change parental functions; alignment quality and retained performance depend on dataset/method. Better than naïve crossover is not necessarily better than parents or mutation-only search.
Source record
- Reference ID
- safe-crossover-neuron-alignment-2020
- Archived version
- arXiv:2003.10306v3, 4 May 2020; GECCO conference manuscript
- Summary provenance
- reference-library/papers/safe-crossover-neuron-alignment-2020/metadata.json · summary
- Review status
- Targeted primary full-text review; not exhaustive reading or reproduction
- Reading scope
- Selected introduction/method passages PDF pp.1–4; complete pp.5–8 on matching, permutation, arithmetic crossover, experiments/results and discussion. Figure 3 visually checked on original p.8. References/code not exhaustively audited; no independent reproduction.