IIRW 2018 information to be available soon.
Training Fully Connected Networks with Non-Volatile Memories: prospects and challenges
Non-Volatile arrays of emerging memories, such as resistive memory (RRAM) or Phase-Change Memory (PCM), enable the acceleration of training and forward inference of Fully Connected (FC) Deep Neural Networks (DNNs), based on the backpropagation algorithm. However, obtained accuracies were unacceptably low due to device non-idealities. Here we show our recent work which, by introducing a novel weight structure composed of two PCMs and two 3-Transistor 1-Capacitor architectures, enables software-equivalent accuracy. Results based on real devices are provided for different datasets (MNIST, MNIST-backrand, CIFAR-10, CIFAR-100). The novel structure based on two pairs of varying significance enables accurate training, also thanks to transfer and polarity inversion techniques, thus paving the way to the development of future technology for Artificial Intelligence based on analog resistive memories.
Stefano Ambrogio obtained his PhD in 2016 in Italy, at Politecnico di Milano, under the supervision of Prof. Daniele Ielmini, working on the reliability of resistive memories and their application on neuromorphic networks. He is now working as a PostDoctoral Researcher at IBM- Research, Almaden, in the Neuromorphic Devices and Architectures Team, working on hardware accelerators based on Non-Volatile Memories for neural networks.