The development of artificial intelligence is progressing at a rapid pace. The next AI generation is already on the advance. And with this generation, we are getting closer and closer to the idea of approximating humans. This is because we are making use of a well-known concept from neuroscience – the function of spikes and neurons.
A new AI generation is upon us
The first generation of artificial intelligence was rule-based imitated logic to draw reasoned conclusions within a specific and bounded problem. This type of application is ideal for monitoring or optimizing a process. The AI generation that followed is mainly concerned with perception and sensing. A Deep Learning Network (DLN) that analyzes the content of a video or image data would be a good example here. The so-called Deep Neural Networks have already arrived in the application by means of classical technologies such as SRAM or Flash-based and initially imitate the parallelism and efficiency of the brain. Further reduction and miniaturization of energy consumption for edge applications is possible through innovative technologies.
The coming next generation of AI extends these capabilities and corresponds with human cognition, such as autonomous adaptation or the ability to interpret. This evolution is crucial to overcome the breaking points of current AI solutions based on neural network training and inference. This is because these, in turn, depend on deterministic and literal assessments of events that often lack context and a universally valid understanding. Thus, the next generation of AI must be able to respond to new situations and abstractions in order to automate everyday human activities. Spiking Neural Networks (SNN) additionally attempts to physically replicate the temporal component of the functionality of neurons and synapses. This allows for greater energy efficiency and plasticity.
One of the definitive challenges in neuromorphic research is to get at human flexibility. The ability to learn from unstructured stimuli while being as energy efficient as the human brain becomes one of the greatest research challenges. The computer building blocks within neuromorphic computing systems are analogous to the logic of human neurons. A Spiking Neural Network (SNNs) is a new model to arrange these elements to be able to mimic human neural networks.
Each neuron in the Spiking Neural Network (SNN) can be triggered independently, and as it does so, it sends pulsing signals to other neurons in the network, which in turn can instantly change their electrical state. By decoding this information within its own signals and on its own timing, SNNs simulate natural learning processes by dynamically mapping synapses between artificial neurons in response to stimuli.
So what can we expect?
Currently, neuromorphic computer systems are still being researched. So far, there is still a lot of development and prototyping going on. The technology is quickly gaining momentum and large corporations, such as Intel or IBM are participating in the promising technology in research projects.
Nevertheless, there are currently still some critical issues that need to be addressed in order for the application of neuromorphic systems to succeed. On the hardware side, we are faced with memory density, the limitation of which is a problem for the creation of robust chips. Especially when it comes to developing chips on which direct learning is to be performed, high-precision synaptic weights must be deposited along with the synapses and the neurons. Here, developments in new integrated circuit components, such as the memristor (a memristor is a neologism of the two terms “memory” and “resistor”) or other nanotechnologies can be supportive.
On the software side, we need to formulate use cases and define relevant problems in which neuromorphic systems can be used to generate enough training data. The problems solved so far by AI technologies are specifically formulated for use cases that do not make use of neuromorphic computing. To illustrate the problem a bit more clearly, let’s take object tracking in computer vision as an example. Object tracking is usually performed by an ordinary camera that records and processes multiple frames of the object. Painterly data from a neuromorphic perspective consists of a continuous stream of spiking signals based on the activation of retinal neurons upon incoming light. The neuromorphic approach gives us a different perspective on how we look at a problem. In order to use this application properly, we need to first look at the problems we want it to solve for us.
There is a lot of AI research going on
One promising project that is addressing these very problems in this area is the Human Brain Project. The initiative, which is co-funded by the European Union, offers a platform for researchers from various fields. The neuromorphic computing platform is designed for computational neuroscience and machine learning. Platform users can analyze network implementations such as simplified versions of brain models provided by HBP Brain Simulation or generic circuit models based on theoretical work. The platform may also be of interest to researchers from industry and technology companies to simulate and test company-specific applications. Compared to other high-performance computing resources, neuromorphic systems offer faster speeds with lower energy consumption.