Lessons from patenting to develop Artificial Intelligence
In my series on “intelligence algorithms” I did not wish to deprive you from this article I wrote a couple of years ago (you can also find it here in an adapted form: http://ezinearticles.com/?The-OWLs-of-Minerva-Only-Fly-at-Dusk---Patently-Intelligent-Ontologies&id=5698710). I also add part of another article I wrote that fits in this topic.
The development of WAGI, Web Artificial General Intelligence, can for instance involve an intelligence algorithm with two metasystem transitions as I explained in my seminal post “Is Intelligence an algorithm?" In his book "Creating Internet Intelligence" Ben Goertzel also implicitly describes this. Steps 3 and 6 I mentioned in my earlier post are the most important steps in that they identify differences, correspondences and spatio-temporal relations are mapped as patterns. A pattern P is said to be grounded in a mind, when the mind contains a number of specific entities, wherein P is in fact a pattern. From correspondences, shared meaning and grounded patterns, abstractions and simplification rules can be derived, whereas differences prompt for the evaluation towards possible modification.
For the abstraction and simplification processes wherein from numerous data events patterns are derived, Artificial Intelligence programs exist and are developed, but they are often dedicated to a very specific niche. When it comes to numerical data, such as in stock market analysis, commercial activity analysis, scientific experimental data etc. or spatiotemporal data such as traffic systems or rule and pattern based data, such as in games, these programs work fairly well for their specific niche. What Goertzel is attempting in the OpenCog software and the Novamente project is bringing these features to general, niche-independent cognition, to the world of Artificial General Intelligence (AGI). Here the data mining which involves a great deal of analysis of a linguistic and semantic nature is of a quite different order.
Although quite a number of programs exist (e.g. DOGMA; OBO, OWL: Web Ontology Language etc.) exist and a lot of work has been done in the field of Ontology (ontology in the field of AI is a "formal, explicit specification of a shared conceptualisation") there is still room for improvement of rules and schemes helping in establishing ontologies.
It is here where the daily work of patent attorneys and patent examiners can provide ideas for development in the field of Ontology and Artificial Intelligence. In fact a great deal of the jobs of patent attorneys and patent examiners involve establishing ontologies. When a patent attorney drafts a claim for an invention, which is a specific entity, he tries to conceptualise in what way the invention can be described in the most general way, whilst maintaining all essential features for defining the invention. Upon drafting an application he has to take into account all possible components of an ontology being:
• Individuals: instances or objects (the basic or "ground level" objects): i.e. the specific entities on which a pattern is grounded, of which at least one must be described in a detailed manner and which can be claimed in dependent claims.
• Classes: sets, collections, concepts, types of objects, or kinds of things: The claim dependency structure, the so-called claim-tree has various kinds of intermediate generalisations before arriving at individual specific entities.
• Attributes: aspects, properties, features, characteristics, or parameters that objects (and classes) can have: A claim most essentially exists of a list of features.
• Relations: ways in which classes and individuals are or can be related to one another. E.g. by means of the dependency in the claim tree.
• Function terms: complex structures formed from certain relations that can be used in place of an individual term in a statement: e.g. the so-called "functional features" which encompass a series of specific entities.
• Restrictions: formally stated descriptions of what must be true in order for some assertion to be accepted as input (e.g. disclaimers, proviso's).
• Rules: statements in the form of an if-then (antecedent-consequent) sentence that describe the logical inferences that can be drawn from an assertion in a particular form: result in dependent claims.
• Axioms: assertions (including rules) in a logical form that together comprise the overall theory the ontology describes in its domain of application. This is most often done in the description; it amounts to giving a plausible explanation of why the structural and functional features give rise to the described technical effect the invention has over the prior art.
• Events: the changing of attributes or relations: which lead to the drafting of different independent claims.
In an astute manner patent attorneys are extremely proficient in this process. With a minimum of generalised features and functional relations between those features, so as to warrant a claim which is as broad as possible without infringing teachings from the prior art, they arrive at giving an ontological definition of an invention.
The whole process of drafting a patent application and especially a successful claim tree depends on the proficiency of the patent attorney to identify classes and sub-classes: hypernyms and hyponyms. In the feature-description he'll have to use holonyms and meronyms, describing wholes and parts, respectively. And in the ideal situation the broadest independent claim has been generalised in such a manner that prima facie it is difficult to see what concrete types of inventions fall under the conceptualisation.
And it doesn't stop there: The differences as regards the prior art prompt for the evaluation towards possible modification and/or additional industrial applications.
When a patent examiner has to evaluate a patent application, he has to go through this process in reverse order. He has to find out which specific entities have allowed for the generalisation and he has to imagine, what existing types of inventions could possibly fall under the scope of the generalised claims. He has to identify which features (structural and/or functional) are responsible for the technical effect over the prior art.
From those notions he can then build a search strategy for identifying relevant prior art, which anticipates and falls within the scope of the claimed subject-matter. For this search strategy to be complete he must combine a set of search concepts which reflect all individual essential features describing the invention. The search will start with some concrete examples of individual entities and synonyms at one level but when simple search strategies fail, the examiner will have to define (in as far as such has not been done by the patent attorney) hypernyms and hyponyms of the features and combine these. Or he'll have to describe a feature as a set of meronyms or conversely a set of features as a holonym.
Nasty problems occur often with acronyms which have more than one meaning, i.e. they are homonyms or polysemous terms, which lead to search hits, which have too many documents. Then the Boolean operator NOT must be added in an additional search statement so as to filter out the irrelevant documents, the so-called "noise".
Antonyms at close distance to negating terms as "not","non","un","dis" or "without" can also lead to positive results. If hits sets contain too many members narrowing down must occur, by adding more search terms or more specific search terms. Additionally, search terms that have a defined relationship can be combined in a specified manner so as to warrant a proximity between the terms: this is done with so-called “proximity operators”, which are more powerful in those instances than simple Boolean "AND" operators. Conversely, if a hit set has too few members, it can be expanded by using more general terms, less search statements or less strict proximities.
In fact in building a search strategy, the search examiner is making a very detailed partial Ontology, and it is a pity (but a logic consequence of the requirement of secrecy) that these ontologies are not stored in a publicly accessible database in analogy the Semantic Web. In addition the community of patent examiners has created and still creates a very detailed classification scheme such as the IPC, which can suitably be used as inspiration in the development of ontological classification schemes. It would also be useful for everybody (not just patent professionals, scientists, inventors and AI-ontology developers) if search engines such as Google and Yahoo would finally make proximity operators available.
There is a lot of criticism from the world of scientists and inventors on the inadequate results that web based search engines deliver (see e.g. Grivell ). The search engines employed by the patent offices are in many respects far superior. Unfortunately for you, they are not accessible to the public. In any way the AI-bot based crawlers and spiders do not go into the deep web databases, where extremely relevant information may be waiting for you.
Ontologies stored in a specific database with links to other deep web databases that are completely searchable in combination with non-spider/non-crawler data mining bots may be a great step forward in information provision.
Kurzweilai is a site about accelerating intelligence. It aims to attain the so-called technological singularity, which will enable mankind to transcend its nature, via an "intelligence explosion". In other words the last invention man needs to make is an artificial intelligence that can improve itself limitlessly.
As a patent examiner I have had ideas about producing this ultimate invention with a slightly different accent. I call this ultimate invention the meta-invention. It is an invention which generates inventions.
Patent examiners evaluate claims (drafted by patent attorneys), which are basically abstracted ontological descriptions delimiting the scope of the invention.
This evaluation process first of all entails assessing novelty by establishing if similar inventions have any differences. If so, the next step is the assessment of inventive step by assessing the obviousness of the invention. In Europe this is done via a set of rules called the problem-solution-approach.
I have had the idea to modify this analysis protocol and turn it into an active invention generator. This requires some (dry) explanations:
Usually the difference over the closest state of the art is analysed and it is evaluated if this difference entails a technical effect. If so the problem to be solved is in the most general sense formulated as "how to modify invention X in such a way as to obtain effect Y".
If there are documents Z from the same or neighbouring technical fields where this effect has been obtained with a similar (yet more different) invention(s) Z and if there is a pointer encouraging to use this solution, whenever the associated type of problem to obtain effect Y is present, then it is concluded that it was obvious to incorporate this solution in invention X.
If you put an important part of this in algorithm form for an AIbot that searches to improve inventions, you create an invention generator (as well as a semantics generator).
The AIbot starts with a given invention X. It searches for problems arising in this type of invention in terms of suboptimal effects or results. Then it starts looking for improved versions of this effect in the same or neighbouring technical fields, in different inventions aiming for a similar purpose.
It evaluates the differences between the documents Z found and X. The most promising document will be the one which is structurally the most similar. If the effect in Z is attributable to certain elements Q missing from X, it will try to modify X so as to incorporate Q.
At first this can be a way to accelerate the generation of ideas for research programs. In the beginning humans will still evaluate whether such a proposition is worthwhile implementing, but after a while if the AI system becomes autonomous enough, it may contribute to achieving the technological singularity.
If this algorithmic system starts to evaluate itself as an invention, it may come up with modifications leading to bolder type of generations of meta-inventions, where "similar purpose" becomes a more hazy definition allowing for finding effects in more dissimilar technical fields, thus combining less related concepts and leading to more break-through type of inventions (as well as generating more trash and noise; yet better systems will eventually learn how to avoid creating noise).
If this algorithmic system starts to evaluate cybernetic systems the outcome may be very interesting, providing a core for self-reflection and a runaway of intelligence improvements.
Key to a successful runaway is that cybernetic systems become hierarchically stratified. Invention implementing subsystems may be increasingly focussed on the perfection of a task, but should not have the flexibility of the "invention generator" itself. The core invention generator should remain as independent as possible and merely propose new solutions without getting involved in the implementation, so as to remain maximally flexible and to obey the adage: "Ideas are more important than execution". However, the downstream technological application of the suggested idea is extremely important –even indispensable- to verify the correctness of the generated hypotheses. This can be carried out by connected implementation systems that will feedback the degree of success to the invention generator, which stores the found solution in an appropriate database. It will be the 15th meta-invention: The invention of meta-generation: Autopoietic self-feedback.
Our technological progress is heading toward an intelligence explosion via artificial intelligence. I hope the suggestions made in this chapter as to how to accelerate this process via a computer program that generates inventions will one day contribute to this technological apotheosis.
Reference:
L.Grivell, EMBO reports 7, pp.10-13, 2006.