Preliminary Discussion - rbjones/HoLoTruth GitHub Wiki

The deductive cloud is a repository for knowledge about logic, mathematics, and about the physical world we inhabit, in suitable form for rigorous deductive reasoning by intelligent artefacts.

To devise an architecture for such a system, to explain that architecture and give grounds for the supposition that the architecture will deliver the desired capabilities, and will be trustworthy, can only be done in the context of a philosophical framework which embraces metaphysics, logic, philosophy of language, epistemology, science and more. The enterprise depends upon, and potentially disrupts, very many branches of philosophy. To devise an architecture which coherently embraces all those things which potentially fall under the scope of our deductive cloud and the synthetic intelligence we seek in it, requires a broad philosophical synthesis. But for a philosophy to provide a stable base for this enterprise it must be simple and solid. I therefore believe that the kind of minimalistic philosophy belonging to the positivistic tradition is likely to serve best, and will seek to articulate such a system.

I see the philosophy as a prerequisite for good architectural thinking, and the architecture itself as a kind of constructive philosophy. Taking cognitive capabilities as a matter for design, demands a perspective on them very different from that traditional in philosophy.

A fundamental issue, and one which has been a matter of controversy throughout the history of artificial intelligence, is the manner in which knowledge is represented. This is particularly significant in that the approach to AI which is at present making the greatest impact is sometimes associated with the idea that no "representation" is needed. I shall not here consider that extreme position, but recognise that the manner of representation can vary very widely, between the symbolic representations which seem to be of the essence in deductive reasoning, and the distributed weightings which we might regard as a kind of representation of knowledge in a neural net.

That the main purpose of this project is the exploitation of formal deductive inference makes the symbolic representation of knowledge a core feature of the project and this will be the focus of much of our attention both in these philosophical preliminaries and in the architectural sketch which follows it. Despite this, the exploitation of deep neural nets will in my opinion be essential to achieving anything close to the intended deductive intelligence, and these nets will store knowledge in ways more similar to synaptic weights than logical formulae.