Image Classification by Evolving Bytecode
a year ago
- #machine learning
- #genetic programming
- #MNIST classification
- Investigates evolving bytecode of a biologically-inspired virtual machine (Zyme) for machine learning.
- Focuses on classifying handwritten digits from a subset of the MNIST dataset.
- Achieves consistent accuracy improvements through random mutations over 50 generations.
- Zyme's virtual machine is designed for evolvability, with a strand-based programming paradigm.
- Initial program performance improves from random guessing (~25%) to up to ~50% accuracy.
- Highlights potential for genetic programming with specialized architectures.
- Discusses limitations and open questions, including scalability and interpretability.
- Suggests future research directions for solving the full MNIST task competitively.