[20a-M206-1~10] FS.1 Focused Session "AI Electronics"

▼ [20a-M206-5] Folded-in-time deep neural networks: Emulating deep neural networks via delay dynamics

〇(P)Andre Roehm^1,2, Florian Stelzer^3,4, Raul Vicente³, Ingo Fischer², Serhiy Yanchuk⁴, Satoshi Sunada⁵, Tomoaki Niiyama⁵, Ryoichi Horisaki¹, Makoto Naruse¹ (1.Univ. of Tokyo, 2.IFISC, CSIC-UIB, 3.Univ. of Tartu, 4.TU Berlin, 5.Kanazawa Univ.)

Keywords:neural network, delay dynamics, adjoint control

Deep neural networks (DNN) can successfully perform a large variety of tasks through training, but are increasingly consuming more energy and larger computers to run. New fundamental ways to implement information processing with physical systems can open up alternatives for different applications. In this presentation, recent advances on how a deep neural network can be emulated by a delay-system are shown. We present numerical results on MNIST, CIFAR-10 and other image recognition tasks via the use of a Folded-in-time Deep Neural Network (Fit-DNN). The hardware requirements for a Fit-DNN would be much lower than for conventional physical implementations of DNNs.