Genetic Programming with Memory for Approximate Data Reconstruction

book chapter

English

This chapter addresses the computation-memorization trade-offs in the context of genetic programming (GP). We introduce genetic programming with memory (GPM) in which GP evolves not only the expression but also the content of a small local memory to better approximate the original data set. In particular, we evolved expression-memory pairs that can serve as weight generators and thus approximate the weights associated with convolutional layers of some convolutional neural networks (CNNs). This is potentially interesting for the efficient implementations of hardware accelerators of CNNs in which memory access is significantly more energy-demanding than arithmetic operations. In our approach, most of the weights are approximated using an evolved expression; only some fraction of them must be read from memory. For example, if memory contains 10% of the original weights, the weight generator evolved for a convolutional layer can approximate the original weights such that the CNN utilizing the generated weights shows less than a 1% drop in the classification accuracy on the MNIST data set. The memory requirements are reduced 3.1x or 12.6x for 8-bit or 32-bit weights, respectively. Additional experiments conducted for more complex CNNs and challenging image classification benchmarks show various impacts of weights' approximation on classification accuracy.

genetic programming, convolutional neural network, approximate computing, hardware accelerator, classification, energy

SEKANINA, L.; JŮZA, T.

28. 2. 2025

Springer Nature Singapore

Singapore

978-981-9600-76-2

Genetic Programming Theory and Practice XXI

199

218

20

https://link.springer.com/chapter/10.1007/978-981-96-0077-9_10

GPTP_XXI_SekaninaJuza_web.pdf