Artificial Intelligence + Hardware
New publication by Francisco Ortega, School of Mathematical Sciences and Information Technology
The latest scientific publicatión by Francisco Ortega1 2, José Jerez2, Iván Gómez2 and Leonardo Franco2 was released on the scientific journal Integrated Computer-Aided Engineering. It was published on january 13th and describes the implementation of hardware for Artificial Intelligence scheme hidden learning process.
To extract information from a data cluster, large scale artificial neural networks with multilayer architecture need long computational times or high performance computational solutions, like supercomputers. On this paper, Francisco and his colleagues, proposed the implementation of FPGA hardware for deep learning, as an alternative method.
A neural network is an artificial intelligence scheme designed to mimic the human brain. There are simple systems, like a neuron tree, and very large scale systems comparable to a human brain. Large scale neural networks are highly used for Deep Learning systems in artificial intelligence processes, extracting information of the network data. This term refers to artificial intelligence systems that learn from the data they collect. A well known example of this is AlphaGo made by the company DeepMind.
Francisco Ortega’s team inserted an artificial intelligence system based on a deep learning algorithm to an FPGA (Field Programmable Gate Array) using layer multiplexing to simulate a leveled data architecture, but precising the resources of a one layer hardware. They simulated 127 hidden layers with a maximum number of 60 “neurons” on each layer. This allows the system of neural networks to gather an important amount of data without exerting pressure over a processor that can’t perform tasks parallely.
The research group confirmed the correct implementation of the algorithm and compared computational times of deep learning processes with FPGA to those obtained from a multicore supercomputer and observed a clear advantage on the proposed system. This is an important step towards the FPGA implementation of deep neural networks, one of the most novel and successful existing models for prediction problems that can go from device controlling to beating the world champion of Go.