Voronin I., Segeda N., Kaliberda N.
Oles Honchar Dnipropetrovsk National University
THE COMPUTER THAT NEVER CRASHES
OUT of chaos, comes order. A computer that mimics the apparent randomness found in nature can instantly recover from crashes by repairing corrupted data.
It seems to be fantastic for many people, but University College London (UCL) proves that systemic computers could allow drones to reprogram themselves to cope with combat damage, or help create more realistic models of the human brain.
Everyday computers are ill suited to modelling natural processes such as how neurons work or how bees swarm. This is because they plod along sequentially, executing one instruction at a time. "Nature isn't like that," says UCL computer scientist Peter Bentley. "Its processes are distributed, decentralised and probabilistic. And they are fault tolerant, able to heal themselves. A computer should be able to do that."
Systemic computation adopts a holistic analysis approach of systems embracing the significant importance of the interactions of their fundamental elements and their environment. Its intention is to resemble natural computation, in order to simulate biological processes effectively, by following a set of simple conventions:
1) everything is a system,
2) systems may comprise or share other nested systems,
3) systems can be transformed but never destroyed or created from nothing,
4) interaction between systems may cause transformation of those systems according to a contextual system,
5) all systems can potentially act as context and interact in some context,
6) the transformation of systems is constrained by the scope of systems, and finally,
7) computation is transformation.
It doesn't sound like it should work, but it does. Bentley will tell a conference on evolvable systems in Singapore in April that it works much faster than expected.
The interaction of two systems can be described by the systems themselves and a third “contextual” system (which is referred to as context) which denotes how/if the interacting systems are transformed after their interaction. The notions of schemata and transformation function are used to describe the interaction. Each system comprises of three parts, two schemata and one transformation function (Fig. 1). The function consists of an instruction from the SC instruction set (more advanced SC implementations may allow a transformation function to comprise multiple instructions). Both systems may change after an interaction, which implies circular causality (each system may affect the other). The scope here, as in nature, is an important factor. The scope of a system defines the neighborhood (which can be other than spatial) in which the system can interact with other systems in a certain way, denoted by the context. Systems are represented as binary strings.
Pairs of systems always interact with a context; these systems constitute a valid triplet. The schemata of the context provide templates for the operand systems to match in order to interact, provided that all three systems belong in the same scope. Thus all computations involve:
• finding valid triplets (context and two matching systems in a shared scope) and
• updating the two systems according to the transformation function in the context.
Fig. 1. SC notation and systems representation:
a) graphical representation of a system; b) the three elements of a system
The pair are now working on teaching the computer to rewrite its own code in response to changes in its environment, through machine learning.
"It's interesting work," says Steve Furber at the University of Manchester, UK, who is developing a billion-neuron, brain-like computer called Spinnaker (see "Build yourself a brain"). Indeed, he could even help out the UCL team. "Spinnaker would be a good programmable platform for modelling much larger-scale systemic computing systems," he says.