Neuromorphic Computing - Future of AI?

The human brain is freaky complex, yet energy-efficient cognitive system. Artificial Intelligence evidently inspired by the way biological brain functions. Some of the AI Neural network concepts like CNN are comparable to the brain. These networks create artificial nodes instead of neurons.

Neuromorphic Computing makes use of engineering, biological, electronics, computer science, physics and neuroscience theories to create an artificial neural system that's similar to the architecture of the biological brain. Neuromorphic computing is also called as Neuromorphic Engineering.

To understand Neuromorphic Engineering we need to understand first how the human brain works. The human brain consists of billions of cells within the nervous system called neurons. These neurons communicate with each other in a unique encoded signal. You can watch the below video to understand how the neurons in the brain work.

Neuromorphic is around for a while and companies like Intel, Qualcomm is trying to develop a dedicated chip which can utilize less power and do more complex calculations. The idea of neuromorphic chips dates back decades. Rise of mobile computing and miniature of digital devices pushing these companies to innovate the underlying silicon technology for the next generation.

"One of the most appealing attributes of these neural networks is their portability to low-power neuromorphic hardware," reads a September 2018 IBM neuromorphic patent application.

As Moore's Law pushed chip designers to pack more transistors onto circuits, the number of interconnections between those transistors multiplied over and over again. Even the no. of cores are increased in day to day computing, the chips are not used as efficient it can be. One clock cycle cannot communicate between all the logic transistors in these chips. Alternatively, Neuromorphic computing is achieved by creating a "Neuristor" circuit which behaves in the same way how the neurons in the brain activity.

According to one of the research article,

A vital component of this neuristor circuit was created using niobium dioxide (NbO2), which replicates the switching behavior observed in ion channels within biological neurons. These NbO2 devices are created by applying a large voltage across a non-conductive niobium pentoxide (Nb2O5) film, causing the formation of conductive NbO2 filaments which are responsible for the important switching behavior. Unfortunately, this high-voltage and time-consuming post-fabrication process make it nearly impossible to create the dense circuits needed for complex computer processors. In addition, these NbO2 devices require an additional companion capacitor to function properly within the neuristor circuit, making them more complex and unwieldy to implement.

Machine Learning/Deep Learning is performed with the prefabricated chips which cannot form connections where there have been no prior connections. Neuromorphic chips, in contrast, will be able to create connections between the neurons on demand basis.

To start with the basic concept of how the DNN can work with these chips, researchers observed that in a brain not all neurons are activated every time. Neurons send selective signals across the chain and the signals are encoded in the someway. These signals are nothing but the train of spikes and thus the research is in the direction to understand if the data is encoded in the amplitude, frequency or the latency between these signal spikes.

In traditional deep neural nets, all the neurons are activated each time as per the single simple activation function. Since not all the neurons are activated at every time, one neuron network in the neuromorphic chip can replace 100's in the traditional neural network yielding more efficiency.

As far as the current silicon technology goes, it is apparent that the traditional chips cannot be used for future computing which involves billions of calculations per second. Even if the neuromorphic chips are nowhere near the efficiency of the brain, it can outperform the current computing technology.

It will be interesting to see if these chips can create a robot which act, think and work as a human in the near future. What if these chips are able to connect to the human brain in any case?. Welcome to the future!