What do I want from AI? Design of protein compounds, computation of elemental reactions and processes in atmosphere are very interesting tasks, of course. But it is far from me. I would like AI to understand me rather than simply identify words and fulfil voice commands. Though I should be thankful for the latter. But what I really want is that it could take hints. I want to be understood like in “Her”.
But how could we make a machine, operating just “0” and “1”, act like that? The answer seems rather simple — we should reverse-engineer human mental processes and patterns of information storage and handling. In other words, we should develop a complete model of human psyche.
After we had already started to develop strong AI, we found some works by Alan Turing (1), Marvin Minsky (2) and Ray Kurzweil (3). Several concepts we found there totally agree with our ideas. Only if AI is like us, it will understand us.
A few sentences about our background. We are a team of psychologists, psychoanalysts and mathematicians. Since 2008, we have specialized on modelling of human mental processes and realized our progress in specific solutions. Additionally, we have revised all psychological theories. Using our view on mentality of Internet users we developed solutions that recommended news articles, considered influence of personal traits on search behavior, and targeted consumers of specific goods based on Internet activity. Those cases showed greater performance that their equivalents in the market.
One of the main traits of our approach is successive realization of maturing mind and its cognitive processes. We gradually supplement new functions,- Piaget’s transformation process (4),- emerging on each developmental stage. We enriched Piaget’s fundamental concepts with some elements of psychoanalytical theories (5). Also we applied some phenomenological methods to study age-related psychological structures and processes (6).
At the very beginning we had to outline some conditions because of resource limitation. We did not involve graphics and executive mechanisms at that phase. Solutions based on neural networks showed a good performance at both processes. Nevertheless, our focus on higher nervous activity admits integration of such solutions.
Also our approach related to human needs is worth mentioning. We suppose that it is inappropriate to realize all human needs as AI’s needs. Some of them relate to sensors directly (e.g. sexual need), some others — indirectly (e.g. safety need). That is why one of our know-hows is AI’s need structure that conditions, from one side, normal functioning, and, from the other side, development of knowledge about outer world in a humanlike way. Further, we will talk about needs in AI structure. But what should be mentioned here is that they are a part of semantic network, some specific part, and they determine AI functioning as a system (e.g. task priority).
AI structure itself.
All information is stored as a semantic network. Its nodes are any phenomena which human mind can determine. Usually, they are objects, actions, traits, abstract concepts. As for interconnections, up to now there are several solutions tending to copy human knowledge structure and operating from 5 to 30 TYPES of interconnection. We apply only those that actually take place at the corresponding age. As result, the semantic net re-creating cognitive structure of a three-year old child contains 4 types of interconnection; and such of seven-year old child — 15 types.
Z type interconnection reflects particular identifier. As soon as there is the “red-squared cloth” in the perceptual field then AI makes conclusion that this is “my sweater”.
W type reflects characteristics. It enables to process information involving characteristics and classify phenomena according to the available ones.
U type transfers all characteristics from the superordinate node to the subordinate one(s). “My sweater” receives all characteristics from “dress” — “wear”, “put on”, “get dirty”, etc.
I would like to mention again — we re-create some average age-related psyche.
When developing our semantic net we found several interesting things:
1. Semantic net has some focuses — elements that condition variance in interconnection allocation. Many chains involving these interconnections are focus-oriented. These focuses are needs.
2. Transformation processes touch upon knowledge structure. Each successive developmental stage requires changes in interconnection structure.
3. While creating a standard net we were constantly running across personal differentiations. And most of them were typical. In other words, number of variants is limited; it is conditioned by either personal traits, or environment. These are psychological types.
One other part of our AI includes algorithms putting data into semantic net or operating data of semantic net. We rely on speech structure to realize understanding. Our algorithm detects connections among elements of declarative sentences that were previously processed by the Stanford natural language parser to process; then all connections are checked up to be present in semantic net, and the new ones are added.
Algorithms operating data of semantic net depends on a particular task. They take into account both characteristics of nodes (e.g. part of speech) and characteristics of interconnection.
At the moment our system can understand and assimilate information from the limited number of subject fields usually introduced to a seven-year old child but without any limits within those subject fields.
Several notions about structural complicacy related to maturing.
The first problem we faced was how to form prior verbal structures. Psychoanalysis offered a list permitting to form some nodes and interconnections but that list was far from being completed. So we formed only those nodes that were parts of structures widely used in later age stages. That was nodes related to such phenomena as “satisfaction”, “Self”, “bad”, etc. At that stage we also developed structures compensating lack of visual, audial and kinesthetic sensors, but enabling to operate information with corresponding characteristics. In other words, AI does not process video but it is introduced to colors, shape, etc. as some characteristics of objects.
Verbal stage provided no difficulties — we just developed several algorithms realizing detection of limited number of connection types (namely, three-year old child — 4 types). Demo of such a system (video). Starting from this stage we had a good opportunity to massively compare our system with a standard — real children.
The most interesting things started at the stage for age 5+.
We provided understanding of almost entire spoken language at this stage. The only limitation touched upon multi-level complex sentences.
Because of impossibility to form abstract notions by himself (this transformation process takes place from 8 to 14), children usually used such notions provided by adults. We taught our system for some abstract concepts of the corresponding subject fields so that our semantic net would properly reflect information. For example, we got AI learnt such notions like “alive”, “predator”, “plant”, so that it could understand texts about animals like a child aged 7. That procedure took place in two ways — either an expert composed a text and introduced it to the system, or an expert himself formed necessary nodes and interconnection in DB. Here is a sample text for AI to learn the notion “fish”:
Fish is an animal. Fish has elongated body. Fish has flattened body. Fish has head, jaws, gills, tail, silver scales. Fish lives in water. Fish can swim, sleep, eat, feel pain, fear. Fish does not speak. If fish is out of water then it dies. Fish uses gills for breathing. Fish uses fins for swimming. Salmon, trout, shark are fish. Cat, bear, coyote, alligator, seal, pelican hunt fish.
To realize question answering we had to take into account all possible forms questions could be framed. Up to age 7, children operate pieces of information in a way of scripts though they admit some variance in phrasing. Later answers are composed from separate elements.
For example, if you ask a preschooler “what is bigger — an elephant or a tiger?”, he will easily answer. But if you ask “what is bigger — an elephant or a giraffe?”, it is well possible that he will be puzzled. After age 10, he will answer without any difficulty being able to operate different characteristics related to notion “more” (height, weight, etc.).
At the moment semantic net contains 10 thou nodes and 40 thou interconnections. Our solution is able to understand simple texts.For example, many articles from Simple English Wikipedia can be understood by the system. Now our progress goes slow as far as we ran out of resources. This resulted that our scope of work was realized in code partially.
After we adapt algorithms to comply with 12-year old person and to learn necessary volume of knowledge, we plan to pass the Turing Test. Our opinion is that AI of such complexity manages to simulate adult communication in a chat room.
Next stage suggests development of self-learning algorithms. In our case self-learning algorithms should enable AI to form new types of interconnection for its semantic net, to alter existing algorithms and to develop the new ones to solve different tasks. We almost completed reproduction of psychological processes and their formalization related to this stage. And the problem to compensate difference between semantic net and human knowledge requires some break down. To keep up understanding, it is necessary that AI would form such new types of interconnection that a man forms in the similar situation.
The existing level of our technology enables to develop advanced solution on its base. That is why we invite other teams for partnership. From our side we provide technological base for products or platforms. From partner side we expect expert evaluation of products and resources for realization in program code. Also we invite developers who are keen on ideas mentioned above.
- Computing Machinery and Intelligence. Alan M. Turing, 1950
2. The Society of Mind. New York: Simon & Schuster. Marvin L. Minsky, 1986.
3. How to Create a Mind: The Secret of Human Thought Revealed. Raymond Kurzweil, 2012.
4. The Language and Thought of the Child. Jean Piaget, 1926
5. The Psychoanalysis of Children. Melanie Klein
6. Logical Investigations. Edmund Husserl, 1900–1901