Common basic knowledge for artificial intelligence algorithms

Completed on 05-Jul-2015 (19 days)

Project 6 in full-screenProject 6 in PDF
Introduction >

Artificial intelligence is gradually becoming the defining feature in a big proportion of new products. Just this fact is enough to support the current project, where I am proposing a comprehensive framework aiming to synchronise the huge variety of AI attempts. Note that, equivalently to what has been happening in most of relevant (and complex enough) break-troughs, the advancements in this field have mainly resulted from isolated (or, at least, tailored to solve narrowly-delimited problems) contributions, which were rarely in a position to apply ideas on the lines of "let our work be the continuation of that previous attempt".

Furthermore, some weeks ago I watched a video which can be considered as the ultimate responsible for this project. It was a documentary about IBM's Watson; a rough overview of its development, the intermediate tests and its final triumph against the best Jeopardy! players. Bear in mind that the target performance of this system (i.e., playing Jeopardy! as well as an experienced person in a completely autonomous way) is so demanding that its development can easily be taken as an absolute reference of nowadays' best way to face any situation involving (English) language understanding.

In that video, there was a specific part which I found particularly interesting. The narrator was describing how the developers realised about the need of relying on machine learning in detriment of more casuistic approaches (i.e., huge lists including all the possible alternatives for each case), on account of the big complexity of the analysed problems. I found the holy-grailish way in which the video transmitted these ideas very compatible with the last public-domain trends on this front; but also, and equivalently to what happens with most of public-domain trends on specialised issues, highly inaccurate. That is: even in case of understanding "machine learning" as a well-defined approach accurately outputting the expected results (what is something very far away from the reality), suitable inputs would have to be provided and AI algorithms can only take as inputs raw data describing the modelled behaviour (human communication, in this case) in the aforementioned casuistic form. Thus, by associating this bipartition with the human learning process, "machine learning" would be the equivalent to the memorising/understanding/adaptation/etc. capabilities and "casuistic hardcoding" to all the knowledge received by the given person (e.g., sensorial information, basic concepts, formal education and so on).

The main goal of this project is proposing a comprehensive methodology to facilitate the synchronisation of the aforementioned "basic inputs" among all the AI-based approaches; or, by relying on the closest human-learning analogy: a preliminary proposal for the creation of a global education plan, which might allow any AI approach to learn from others/share its knowledge.