MLMU KE: Neural network implementation: one's own VS a framework? – Rudolf Jaksa

Pozývame Vás na ďalšie MLMU v Košiciach. Prednášať bude Rudolf Jaksa.

Abstract:

A naive implementation of neural networks in the C language can be tens, or hundreds of times faster than the usage of common neural networks frameworks. The simplicity of neural networks learning algorithms goes well with the automatic vectorization in recent versions of GNU C Compiler and in the result native builds are much faster than generic, and newer CPUs are significantly faster than just a few years old ones. We will show possible gains. Concrete speedup depends on the topology of the network and on the CPU itself. To measure it, we prefer the classic Connection updates per second (CUPS) training speed metric, which is more universal and more practical than task-specific benchmarks. “To measure” is important, as the interplay of topology, framework and hardware is complex. Our C code for neural networks training and support tools for automation of the CUPS curves measurement are on GitHub.

Speaker:

Rudo is the Head of AI in Matsuko, working on custom 3D convolutional
architectures for holographic communication. Before he worked on NN predictors and RL for E-commerce in Exponea, and on the industrial and weather NN prediction in Kybernetes/MDJ. For 20 years he taught Neural Networks on the TU Kosice. He worked with Interactive Evolutionary Computation on KID Fukuoka.

Archív minulých meetupov a ďalšie informácie o MLMU nájdete na našej stránke mlmu.sk