Equivalent-accuracy accelerated neural-network training using analogue memoryStefano AmbrogioPritish Narayananet al.2018Nature