AI - what next: cognitive architectures

One of the major questions the AI community raises is: What next? Though accuracy still remains a priority, this problem is already not as acute as it was ten years earlier.?Thanks to the recent breakthroughs in AI, ML, Deep Learning, and development in machine hardware many old problems, such as Computer Vision and NLP have been successfully solved.

AI community is now interested in the solution of multiple problems, such as for instance;

-?????????Interpretation of AI: how to explain and justify the outcomes so they can be trusted.

-?????????Meaningful?labeling of data sets, and transfer from manual to automated labeling

-?????????Auto-labeling, a.k.a self-supervised learning

-?????????Zero-learning, or better say low-effort learning of new cases

-?????????Usage of context information for improving?accuracy??

?In the practical areas, the focus has now shifted to implementation and integration. That requires new approaches rather than just improving accuracy, which is already sufficient for the integration into the practical systems.??It also refers to the discovery and implementation of new use cases for AI.???

So where to look for the new market applications? Following the trend that successfully created modern AI, we can look at what problems aren’t solved yet.?When we create new solutions for a new use case, we may need to understand the strengths and weaknesses of different technologies and find what solutions may complement each other to form a viable system.???????????

To understand all of this in clear terms, we may need to understand a contextual reference system of relevant concepts.?They are:?

·????????certainty and uncertainty;?crisp and fuzzy; conscious and unconscious

·????????perception, learning, and recall; intelligence and memorization

·????????concepts, symbols, and alphabets (or sets);?systems, order, and logic

·????????cognition, inference, synthesis, and decision

Although all these concepts look unrelated, they are linked together in a very systematic way.

There are two levels of information processing in the brain: conscious and unconscious.

There are two different types of information models as well: uncertainty and certainty. Uncertainty is usually represented in the information models with some distributed or fuzzy energy value that describes our confidence. Certainty is best described as high peak confidence energy, distributed across a very small area. Ideally, this is a point with unlimited energy.?

Models and processes may be significantly different for each level. However, a transfer mechanism exists, which allows for the interaction between levels, and one level may not effectively work without the other. Transfer from uncertainty to certainty occurs when making a decision. Where a decision is made consciously or unconsciously, the outcome is always certain, crisp, and conscious.

A lot of information is processed at the unconscious level, and this is usually uncertain or fuzzy information. We cannot explain why and how we got to this conclusion, and such processing looks to us mostly like a recall. Though to get to this state, we need to undergo some training, accumulate some experience and develop a mechanism that may allow us to make such recalls.??

Models based on certainty are known as conscious. Those models are responsible for understanding. They can be described in terms of human language and logic can be used for drawing conclusions. They are also known as semantic models and they have been around for a while. What were the issues with such models in the past that limited their practical application?

In past, the major issue was their representation, where such models were based on linguistic constructs like rules and facts in some simplified human language. But a lot of conscious crisp information goes beyond the human language. For instance, diagrams are a significant part of every design. Diagrams can be very well and clearly understood and described on the conscious level. And it is well said that one diagram may cost thousands of words

The nature of intelligent models is not purely semantic but rather semiotic. And such models in the general cases can be better represented as graphs rather than constructs of human language. Language is just a means of knowledge transfer, and it covers best mostly aspects of serialization of knowledge for the transfer between the individuals.?

Purely linguistic representation hides diagrammatic and semiotic aspects of knowledge models, and that creates obstacles to implementing intelligent operations like the structural analogy, blending, and synthesis that can be understood and implemented as graph transformations.??

A supervised neural algorithm creates a mapping between the uncertain input information and a predefined set of distinctive labels on its output. This is exactly the mechanism of decision that converts uncertainty to certainty.?

Labels are usually implemented as some string of letters that denote words of human language. Such words denote concepts, which are the components of the human cognitive system that describes our knowledge on the conscious level. Concepts and their relations form sets and systems, and concepts are nodes in the conscious conceptual network models.

Human words are not the only way to label a node in conceptual networks. The meaning of the concept in such graphs is determined by its relation to other concepts. Therefore, the actual meaning of the label is defined by the relative position of its node in the conceptual network. In the cognitive network a node can be labeled with any distinctive symbol instead of the word of human language, and it would take the same role as the corresponding word.

A cognitive network fully labeled with linguistic labels is well known as a semantic network. But words are required for understanding by a human. Computers can also use for their “understanding” a node in the relational conceptual or cognitive network. Then they can utilize its relational meaning in the application in the same way as people would utilize the meaning of the words in their cognitive system.

This is a known fact that human languages don’t always have the same sets of colors. Assume for simplicity that we have a neural network that is trained to recognize the following set of standard colors in some language, for instance {Red, Yellow, Blue, Violet}. Assume that the words Orange, Green, or Magenta do not exist in that language.?

Such colors are naturally ordered according to the scale of light waves.?And the order of nodes and symbols in the conceptual network may look like Red-Yellow-Blue-Violet. But a lot of times this network brings a mixed response like Red-Yellow, Yellow-Blue, or Blue-Violet. When such mixed responses create significant clusters, it becomes obvious that the alphabet of symbols- labels, which this network was trained to recognize, is insufficient.

In such a case, we may need to introduce a few more concepts that correspond to the new clusters. We still can use labels Red-Yellow, Yellow-Blue, or Blue-Violet. Or, invent new words like Orange, Green, or Magenta. So we will have our new conceptual network as Red-Orange-Yellow-Green-Blue-Magenta-Violet and corresponding set-alphabet of new color concepts-symbols {Red, Orange, Yellow, Green, Blue, Magenta, Violet }

Naming is not so important. A major thing is that now our conceptual network consists of 7 nodes instead of 4, and we will have a new alphabet with 7 distinctive symbols. This allows us to understand the idea of implicit labeling.?????????????????

There are no explicit symbols or labels in the brain. How does the brain handle this situation? Node or concept in the brain looks like a strong activation response from some local cortical neural assembly. This assembly is connected with other similar local cortical assemblies via remote links. And the remote connection between the isolated remote assemblies that were activated simultaneously or within a short period of time forms a permanent stable link = new relation.

This is the way how the brain creates and maintains cognitive networks. This entire cognitive system may be treated as a consistent set of graphs, similar to the conceptual networks. The labels may come from the connections of outputs of perceptual networks in the cortical areas of lower levels. Such areas were found in the ventral pathway, and neighbor neurons in such areas allow for the recognition of different objects belonging to the same category.

Neurons in those perceptual areas “make decisions” which concepts are presented in the input information and pass those decisions to the upper levels. Given that the previous decision invokes a node responsible for a particular concept, all further processing is done already on the conceptual level. There are also multiple feedback connections. And cognition may affect perception also, which is a known phenomenon in cognitive science.?????

The most interesting thing about such cognitive networks is that they are able to perform intelligent operations, such as synthesis, analysis, and logic. While the discovery of actual neural mechanisms that can do it is still in process, they can be described and implemented as graph algorithms.

This takes off the major burden from the practitioners. Similarly, there is no need to emulate a set of differential equations in the semiconductors where Boolean logic may be sufficient for the description of what the CPU does on the informational level.???????

From this holistic position, we may now look at the problems of labeling. We should look at the labels not as on the isolated sets of words of human language. But we should consider labels rather as nodes of the large consistent conceptual or cognitive networks.

Semantic labels denoted by human words are better understandable for humans. However, in many cases such labeling is impossible. But the semiotic approach for labeling does not require words of the human language and can be used in replacement. Cognitive or conceptual graphs may self-develop, and dynamically create a new node upon its relative locations to other existing labeled nodes. Any distinctive symbol or their combination can label this new node, and this will be sufficient for using it for semi-supervised or self-supervised learning.

Self-supervised learning is needed for the successful creation of autonomous robotic systems and self-driving vehicles. Modern AI is very good in the area of machine learning and perception. But can we just learn everything and always get the needed response to any situation? What if the solution cannot be learned from experience by trial and error but must be created or derived from the context?

Animals can learn pretty well, but creating something that cannot be just learned is beyond their capabilities. Learning and recall don’t mean intelligence.?Intelligence is the ability to consciously derive new complex decisions based on the context that is already known. That requires logical inference, and animals do not possess such capabilities.

Allowing driving someone who isn’t smart enough to predict and handle unknown situations would be extremely dangerous and can jeopardize the whole idea of autonomous driving. For that reason, there is a great interest in combining machine learning with semantic reasoning. And one of the major questions is how those technologies can complement each other.????

Knowledge Graphs are already used in a large number of applications. But their ability to carry out cognitive processes and reasoning has been limited. One of the big problems with modern Knowledge Graphs is that it is difficult and time-consuming to capture semantic knowledge.

But the major issue is that it may not be possible to hardcode everything just in the form of semantic rules and facts.

A more flexible representation is needed, which may allow for coding of diagrammatic knowledge, explicit operations on the structural components, implicit labeling, and dynamic transformation of the knowledge graph for deriving new concepts and solutions on the fly. That will allow modeling intelligent cognitive processes similar to the ones in the human brain.?????????

Cognition doesn’t worth much without perception. Machine Learning can suggest new nodes and connections in the cognitive graphs; find correspondence between nodes and parts of the graph, etc.?Perception is the area, that modern AI can already cover very effectively. Especially if it is complemented with the cognitive system that will be based on the graphs and semiotics. And this will make AI outcomes self-explainable and transparent.

?I believe that this powerful combination of ML, Cognitive Graphs, and Semiotics is going to be the next big breakthrough in AI, because it holds the answers for the next big advancement in the artificial intelligence industry and related fields and markets.????

GLOSSARY

Cognitive network - relational conceptual network – dynamic graph model where nodes represent concepts and their position in the network is relative to other concepts. Graph representation allows for the separation of structural and semantic (or semiotic) components. This allows the implementation of intelligent operations like an analogy, conceptual blending, and synthesis as graph algorithms and transformations in the cognitive network. Also, such graphs are capable of capturing not just linguistic but also diagrammatic knowledge, and in a major sense any systematic knowledge model or process.

Concept – node in the cognitive network. It may represent any qualitative object, entity, pattern, property, process, set, class, category, system, situation, and decision – anything that can be considered as a concept. The concept is always crisp, certain, and conscious in the system.

Label - concept node can be either explicitly labeled with some distinctive symbol or combination of symbols, or implicitly labeled via links to other nodes. Explicit labels are needed for the serialization of knowledge and logical operations on the concepts.????????

Semantic or linguistic labeling – labeling concepts and relations with words or other constructs of human language, which simplifies understanding by human

Semiotic labeling – labeling the same distinctive concepts or relations with the same distinctive symbol or combination of symbols. We can think of such semiotic networks as extensions of semantic networks. While the latter allows for capturing mostly linguistic knowledge, the former allows for capturing diagrammatic or any systematic knowledge as well.?Though such labels may not be easily understood by humans, they can serve the same purpose as an internal language for computers, especially when semantic labeling is not applicable to the concepts?????

Relation – a link between the two concepts, which can be labeled or associated with some relational concept and can serve as a predicate in the logical chain. Two linked concepts are neighbors even if they are remote from each other.

Cognitive processes – set of processes in the cognitive network that allows for implementing logical and structural transformations which change structure and concepts in the cognitive network

Dynamic concept creation – dynamic creation of a new node in the cognitive network.

Dynamic relation creation – the creation of a new link and assigning a logical predicate to the relation. For instance, if predicates allow for a logical chain with a few hops from node A to node D, a new link can be established between A and D, and A can be considered cause and D result. Next time we can skip all intermediary steps and directly get from A to D.

The problem of labeling in AI – assigning meaningful linguistic values to the expected outcomes of machine learning.?Semantic labels are applicable only to supervised learning. However, any other learning types like Semi-Supervised, Self-Supervised, and Unsupervised require a different approach. For instance, matching toward a cognitive network where concepts for labels can be created dynamically as new nodes.

Cognition - the mental process involved in knowing, learning, and understanding things. Synonyms: reasoning, understanding, intelligence. Cognitive models and processes may be represented and expressed in terms of special graph models??

Perception –processes of converting input information in the raw format into meaningful and understandable concepts of a cognitive system. This is equivalent to making a decision about what input information represents. Currently, perception is a well-established domain for AI and ML. Perception and cognition work as a single system and cognition may affect perceptual processes via feedback connections, helping to resolve ambiguities.?

Graph Databases – Commercial technologies that can store, retrieve, and process large graphs

Vertical Markets – markets that can be built upon AI, such as for instance robotics, self-driving vehicles, and autonomous systems.

?????????

Disclaimer: this is not a scientific or industry article but rather some interesting points brought for discussions with peers and colleagues in industry and academia, who might be interested in such topics in their personal and business life.

Thanks,

Gary Kuvich??????????