Advancing evolutionary genetic programming for automated neural architecture discovery. Developing novel mutation operators and fitness landscapes for finding optimal deep learning architectures.
This research project focuses on automating the design of neural network architectures using evolutionary computation techniques. The framework implements sophisticated genetic programming operators that can evolve both the topology and hyperparameters of deep neural networks.
Key innovations include:
- Novel crossover operators that preserve important architectural motifs
- Adaptive mutation strategies based on network performance
- Multi-objective optimization balancing accuracy and computational efficiency
- Parallel evaluation on GPU clusters for faster evolution
The system has been tested on various computer vision tasks, showing consistent improvements over manually designed architectures while requiring significantly less human expertise in network design.