Neuromorphic systems and evolutionary AI

In the field of artificial intelligence, the idea of evolution is a critical one. The very concept of an AI algorithm is one that can grow, change, adapt, and in essence evolve to fit its design requirements. AI research has progressed to such a point that an AI can be trained to achieve a level of performance far surpassing what a human could attain under specific conditions, such as a smart robotics station, offering never before seen precision and speed in smart factories.

However, this high degree of speed and precision does introduce some downsides as well. Training AI to work well in a specific location and task requires a vast amount of data from that location. What’s more, if conditions change, the algorithm may suffer reduced accuracy or cease to function altogether. Evolutionary research and neuromorphic systems is a promising approach for helping AI evolve in a more natural way, through both software and hardware, allowing it to tackle a wider range of conditions without loss of performance.

Creating a robust solution is one of the most important aspects of designing an AI system, especially for industrial applications. AI algorithms are trained to accommodate a wide expected range of conditions based on their expected operation environment, but when incoming data falls outside of the range of training data or becomes too noisy, the system will more than likely fail. Unlike AI, humans can continue to perform tasks even under widely varying conditions, so the goal is to design AI to be more human-like. This is exactly what neuromorphic systems propose to do by seeking to emulate not just the neuron structure of the brain, but the function of the synapses as well through memristors. Synapsis and memristors have the property of changing resistance based on how much current has flowed through them in the past, essentially creating a memory of past values. Thanks to this memory-like functionality the AI decision-making process becomes more robust and less susceptible to failing from less-than-ideal input data.

Going hand in hand with neuromorphic systems is evolutionary research that fundamentally is similar to reinforcement learning but does have some key differences. Reinforcement learning is a common way for AI to be trained by governing a system with a reward function that reinforces correct behaviors and punishes poor ones. This allows the software essentially on its own to evolve a solution that fits the requirements. Where an evolutionary system differs is in the scope of what the reward function manages. Called genetic code, its scope goes far beyond training just software, instead it governs everything from robot design to software to control chip layout. While in a factory the options for robot design are limited, the ability to train robotics control hardware through programmable FPGAs along with the software presents a unique path to improving AI robustness when combined with neuromorphic systems.

Allowing the hardware to be trained alongside the software under these new principles paves the way for an entirely new level of adaptability and robustness in AI. Compared to a traditional machine learning training schema, a neuromorphic system does not need to be trained with only “correct” data. Rather its training process can include non-ideal and highly variable data or even outright failure states. Although it might seem counter-productive to include undesirable outcomes in the training process, this approach allows the system to learn and evolve so it can function accurately even when contending with poor quality data or faulty hardware. More than accounting for failures, a neuromorphic system might also help mitigate the heavy reliance AI has on deployment-specific training data. Rather than resilience against poor quality data, the system would be accounting for environmental differences between the training location and the final deployment site.

The quest for truly human-like AI and robotics still has a long way to go but evolutionary design and neuromorphic systems combining to allow hardware and software to truly evolve side-by-side bring us a step closer to true natural evolution. While this technology still requires substantial development before it is economically viable to replace existing AI implementations in phones and factories, it may see its first applications in the cold reaches of space. The extreme conditions found even in low earth orbit make it difficult for existing AI systems to function accurately, something of paramount importance when operating a million miles from earth. In years to come, neuromorphic systems may become the living mind of the next generation of satellites and deep space exploration crafts pushing the boundary of human understanding.

Siemens Digital Industries Software is driving transformation to enable a digital enterprise where engineering, manufacturing and electronics design meet tomorrow. Xcelerator, the comprehensive and integrated portfolio of software and services from Siemens Digital Industries Software, helps companies of all sizes create and leverage a comprehensive digital twin that provides organizations with new insights, opportunities and levels of automation to drive innovation.

For more information on Siemens Digital Industries Software products and services, visit or follow us on LinkedIn, Twitter, Facebook and Instagram.

Siemens Digital Industries Software – Where today meets tomorrow.

Leave a Reply

This article first appeared on the Siemens Digital Industries Software blog at