Source: https://www.datamation.com/artificial-intelligence/neuromorphic-computing-isnt-path-general-ai/
Neuromorphic computing deals with developing electronic analog circuits to mimic neuro-biological architectures present inside our brains. Any device that uses physical artificial neurons (made from silicon) to carry out the computations can be considered a neuromorphic computer.
Neuromorphic computing models the way the brain works by using the concepts of spiking neural networks. Unlike conventional computing based on transistors turning on or off, one or zero, spiking neural networks can convey information in both the temporal and spatial way like that of the brain and produce more than one of two outputs.
Von Neumann’s architecture, which separates memory and computation, is used in most modern electronics. Von Neumann chips waste time (computations are slowed by the speed of the bus between the computer and memory) and energy because they must shuttle information back and forth between memory and CPU.
Von Neumann architectures will become increasingly challenging to deliver the growth in computing power that we require as time goes on. On the other hand, although deep learning platform development is growing in popularity, those platforms also need an immense amount of computational resources. A new type of non-von Neumann design, known as a neuromorphic architecture, will be required. This is where neuromorphic computing comes in.
You might be thinking about why the human brain is such an appealing model for computing. To clear this out, we can look at the following points.
- Supercomputers can take up several rooms, whereas the brain is compact and can fit in our head.
- Our brain uses about 20 watts which is far less than the energy consumed by the massive supercomputers.
- Supercomputers need cooling systems to keep them from heating up. On the other hand, the brain sits inside our skull at 37°C.
- Harnessing techniques used by our brains could help us to build more powerful computers in the future.
Although the idea of neuromorphic computing seems promising, it does come with a set of challenges that needs to be tackled along the way. The major challenges include:
Massive Computing Resources
The development of neuromorphic computing models requires a massive computational resource. The development is slow and expensive. We need custom hardware in order to run state-of-the-art processes for such development. Also, the software and hardware development for neuromorphic computing must go hand in hand. We can neither develop the algorithms first nor specify the hardware first.
Computing Norms
For so long, data encoding and processing have been based around the von Neumann model, which will need to be reworked for a more common world of neuromorphic computing. For example, our conventional systems take it as a series of individual frames to deal with visual input. In contrast, a neuromorphic processor encodes such information as changes in a visual field over time.
Recreate Programming Languages
Programming languages will also have to be rewritten from the ground up. There are also significant hardware challenges: new generations of memory, storage, and sensor technology will be required to exploit neuromorphic devices fully.
Impact on Hardware and Software Development
Neuromorphic technology could even need a fundamental change in how hardware and software are developed because of the integration between different elements in neuromorphic hardware, such as the integration between memory and processing.
Ethical Concerns
As these systems aim to imitate a human brain, a set of ethical questions may arise. One of the most significant ethical challenges that may be placed on neuromorphic computing is public opinion. Will the public agree with the idea of using robots in the care of their children and elderly parents?
Testing Dilemma
Brains are embodied, and bodies are embedded in an environment. Testing neuromorphic computers often require embedding them in a complex body environment. How do we even prepare such a complex body and environment?
Neuroscience is of very little help when it comes to neuromorphic computing. Computational Neural Models are just high-level representations. They make too many assumptions, have too many parameters. They have no general organizing principle and are (usually) incomprehensible.
Moreover, brains, bodies, and environments are complex systems with large-scale functions that cannot be represented analytically in terms of their lower-level structure and dynamics. The system design approaches are insufficient since the system cannot be broken down into separate components.
All in all, it is difficult to simulate all the details of a biological brain when we don’t even fully understand the function of a simple nervous system. Also, there aren’t many observables to help guide the development of this technology.
