https://www.nature.com/articles/s41467-019-12521-xPublished: 18...

https://www.nature.com/articles/s41467-019-12521-x
Published: 18 October 2019
Building brain-inspired computing.

7. What could be the measure of when the neuromorphic computing is ready to replace the current digital computing?

DS: Ultimately, this will happen when, for instance, the projected performance obtained from the prototypes at small scales show compelling evidence that neuromorphic systems can outperform digital ones and/or be competitive in other metrics at the application level. Judging by the rapidly growing interest from entrepreneurs and many start-ups founded in the recent years, I would say that this might soon happen for neuromorphic inference accelerators. I also heard the possibility that dynamic vision sensor technology, which takes inspiration from biological spiking neural networks, is being close to its commercialization. This, if true, could be another milestone in neuromorphic computing. I would also like to use this opportunity to emphasize the importance of benchmarking when evaluating the potential of analog neuromorphic hardware for deep learning applications. To reach any sensible conclusions, in my opinion, rigorous comparisons of functional and physical performances across different implementations can only be conducted for well-defined applications with clear performance metrics and common benchmarks (such as classification fidelity on MNIST—a handwritten digit image set).

GI: This might happen much sooner than we think, and probably we will not even notice it. In the past whenever a neuromorphic circuit was deployed in a commercial product, it was never advertised or recognized as such (e.g. the Logitech Marble trackball, which used a neuromorphic motion sensor chip to measure the rotation of a ball painted with a random dot pattern, to control the cursor on a screen, see Arreguit, X. et al. IEEE J. Solid State Circ. 31, 1916−1921, 1996). I envisage that soon we will see the first instances of neuromorphic products in low-power intelligent sensors that extract information from the analog signals they measure and make simple (e.g., 1- or 2-bit) decisions without connecting to a cloud. Our cell-phones are already covered with an abundance of sensors (including pressure sensors, accelerometers, etc.), and the need to have ultra-low power always on sensory processing will eventually drive the industry to explore more exotic designs, such as those comprising analog and/or asynchronous digital neuromorphic circuits, and include them in their systems.

JG: I don’t think neuromorphic computing will replace digital computing. Digital computers are great at high precision computing, whilst neuromorphic computing is great for cognitive tasks such as pattern recognition, classification and prediction. Digital and neuromorphic computing will complement each other and future processors will integrate both.