Recently one of the most respected experts in AI processors positioned POLYN’s Neuromorphic Analog Signal Processing (NASP) as a unique technology.  Indeed, neuromorphic designs are novel and have yet to achieve widespread awareness and adoption, even if in-memory computing is already considered mainstream.

The very term “analog” causes misunderstanding, recalling the notion of such aspects of the circuit performance as linearity, noise, input/output impedance, stability, and voltage swings. Speaking of general analog integrated circuits, it is hard to attain the same level of robustness as with digital circuits.

With that said, how indeed does NASP overcome a mismatch of numerous analog neurons?

The answer lies in the term “neuromorphic” as neural networks implemented in a neuromorphic analog circuit. The point is that resilience to errors is a fundamental property of neural networks, and training increases the resilience. 

Circuit non-idealities can be divided into two groups: random and systematic errors.

Systematic errors occur because a typical circuit implementation only approximates an ideal signal processing operation to a limited extent. Such errors are caused, for instance, by the non-linear operating characteristics of analog circuit components or by finite gain of an analog amplifier.

Stochastic (random) errors may happen during the fabrication of integrated circuits and result in a random variation of the parameters of the fabricated on-chip elements.

All these errors, however, can be modeled and addressed during development.

Batch Normalization layers, being building blocks of modern neural networks, significantly reduce network sensitivity to unsystematic errors. Regularization techniques, both standard and custom, also have a beneficial effect. If properly used during neural network training, they may reduce the output dependence on any particular signal of weight, so that a variance in the hardware components does not greatly affect the output signal. POLYN has developed the know-how reducing a neural network sensitivity to errors.

Additional methods are based on hardware design principles: for example, a mismatch between neighboring elements is usually much smaller than the variation of absolute values of the element parameters. Therefore, differential architectures significantly improve precision.

Neuromorphic is the key to success for NASP, recently confirmed by NASP test chip production and validation.

Simply explained, the ability of deep neural networks to handle errors is similar to a fundamental function of error processing by the human nervous system and brain. This fundamental property gives us control over analog device mismatch and allows using such advantages of analog systems as ultra-low power consumption and latency.