Train with noise for accurate machine inference using imperfect hardware

<!--HTML--><p>The machine learning revolution under way brought us neural networks that outperform humans at a variety of tasks ranging from object recognition in images to winning at board games and abstract reasoning. As the field evolves from academic discovery of the capabilities of neural-inspired computations to more widespread deployment of machine agents into the world we inhabit, our attention turns toward cost-effective and energy-efficient hardware capable of performing these kinds of computations with good-enough accuracy and resilience to real-world disturbances.</p> <p>In this talk we will review recent results from Western Digital Research that point the way toward accurate machine inference with hardware that is fundamentally incapable of performing exact computations, such as neuromorphic analog circuits, or even traditional bit storage without the use of error-correcting codes. We will show that, if the network is trained in the presence of the kind of perturbations expected from the imperfect inference hardware, then the accuracy of inference can be almost as good as if exact computations were used throughout.</p> <p><strong>About the speaker</strong></p> <p>Dr. Dejan Vucinic is Director of R&amp;D Engineering at Western Digital Research in San Jose, California. His group has been exploring the impact of emerging non-volatile memories, such as MRAM, ReRAM and PCM, on computer systems architecture. Dejan was a summer intern at the L3 experiment at CERN in 1992, and earned his Ph.D. in experimental particle physics at MIT in 1999 under the direction of Prof. Sphicas.</p>