Original article was published on Artificial Intelligence on Medium
Artificial Cognition in Artificial General Intelligence — Part 1
The Role of Dimensional Variance Phase Shifting (DVPS) Foundation in Cognitive Artificial General Intelligence
Rob Smith, eXacognition
This is one of a series of articles detailing some of the key designs, concepts and thought processes behind the development of an Artificial General Cognition (AGC) which is the key foundation component of Artificial General Intelligence (AGI) systems. In this first article, I will introduce the DVPS foundation that allows our AGC to perceive its world and learn on its own the same way humans do. The balance of the articles will address the architecture behind the building of the ‘general components’ in artificial general intelligence.
One of the most unique elements of human cognition is our ability to anticipate unforeseeable change and use that change to impact our decisions instantaneously. It is unlike anything in the world of common AI development today and is what gives our human intelligence the ability to be ‘general’ or understand, learn and adapt to everything. As complex as it sounds, it is relatively simplistic to achieve once an underlying cognitive foundation exists. Our own human cognition is comprised of a vast foundation on which our learning and knowledge resides and which create the decisions that we make and change throughout our day. It moves as a never ending incremental flow of adjustments to this inherited and changing foundation on which our cognition rests, or our knowledge, by changing, updating and adapting to our cognitive perception with every passing second of every day.
Knowledge in humans is a measured categorization of elements and relationships that is multilayered and multidimensional. It is this way because the categorization is not only temporal, it is contextual, meaning that the same element can have many different classifications that change over time and change according to time. Categorizing the elements of a phrase such as ‘that is a dog’s breakfast’ can have different meaning to each person who receives the phrase through their perception. To you the reader, they are words on a page relative to the context of this article but to someone hearing the phrase, it is relative to the context of which they are in and that is set by both their perception and their knowledge. It is fluid and changing and includes layers and cascades of other elements over a flow of time.
If I say the words ‘that’s a dog’s breakfast’ to you with no other context and then all your perception is immediately turned off, the only grounding for your cognition to work from would exist from your knowledge which is nothing more than a series of memories projected into the future on contextual threads held together by perceptual probabilities of occurrence. An ‘out of the blue’ phrase with no context or perception results in our cognition literally taking a guess at the meaning of the phrase based on our last known perceptive flow and the anticipated extension of that flow into the future along optional cognitive paths. Of all the potential forward anticipated paths available, we choose the one with the highest relative probability to achieve one or more goals within a context of ‘success’. In short, we do what we think is right for the situation by measuring a distance toward one or more goals and we do it instantly.
In advanced artificial general intelligence design, this ability to think as we go is a very complex problem to solve but it is not an insurmountable one. There are a vast array of methods and techniques that we can use to build machines with flowing cognition that constantly improve as time moves forward. This is even more possible if we connect other Artificial General Intelligence systems as if they were nodes on a network and share the knowledge with other machines to expedite the overall improvement in learning of all machines or an overall artificial consciousness. We can do this simply by creating a framework whereby the AGI’s can communicate together from the same base foundation of knowledge creation and cognition. We humans do this from our own base foundation of consistent sensory perception and inherited response intelligence. You and I may speak different languages but in general we sense the exact same thing (except for a small amount of personal variance). An apple all around the world is an apple although it may have a different name or category according to one’s language or context.
Context, however, is critical to this sharing because a piece of fruit is different than a piece of technology like a phone or computer although they may both have the same ‘categorization’ or name. It is our shared method of perception that really binds humans together. Our desire for competition is what stops us from sharing our knowledge completely and yet we do share our knowledge consciously with our children or with others to advance our own personal goals. I share this knowledge because I am relatively old and want a future world to benefit if the information has any value or to spark innovation in a younger individual. My hope is that it helps humans advance to become a violence free group deserving of the title advanced interstellar species.
Our knowledge foundation of shared perception makes us unique as animals in that we encourage in our youth to innovate beyond basic survival actions and that makes humans excel in the game of evolution. However, we are also gifted with knowledge at birth directly as a result of evolution. A baby does not go to school to learn to cry, it just automatically cries after it leaves its mother’s womb. The same thing occurs in animal babies as they instinctively know to get up on their feet or suckle on their mother’s milk after birth. We are gifted with a foundation of knowledge from birth and more importantly, we are gifted with the ability to improve this knowledge right from birth by constantly adapting and changing to a flowing world streaming into our senses.
Contextual Categorization & Relative Importance
Just as we categorize elements of knowledge, humans also categorize context or the relative relationship of knowledge elements to our current perceptions and to other elements & contexts. We do this by linking them together through the use of probabilities. These probabilities define the nature or degree of the relationship to other elements, existing knowledge, perception and even time. A dog is an animal but it is also a ‘friend’ or ‘companion’ and can be a ‘good boi’ one moment and a ‘bad dog’ the next. Not only do the probabilities of relative importance change as the context changes, they are multiple and can ‘cascaded’ to impact the net value of an entire perception. This is a critical part of how we humans understand rapidly changing perceptual context when we are doing something like driving a car and how we are able to so rapidly respond to change that we have no knowledge or awareness of until it occurs. It is also how we dream up things that do not exist like inventions or solutions to problems or creative works.
The interesting part of all of this is that AGI systems do the same thing and can do it faster than a human mind theoretically but only if we humans can build machines that are as massively efficient as a human brain and we can. One Artificial General Cognition design that moves toward this goal involves a staging of our cognition to focus our attention on the perceptions that are most critical to our own self awareness. We do not, or should not, fill our brain with everything we know while driving a car down the road. We should and do focus most of our cognitive resources on the act of driving. It is why using your phone while driving is so dangerous. We humans perform the feat of driving a car safely through traffic with ease by adjusting the relative value of a subset of information within our knowledge to the context of the act of driving a car. This helps us preserve and focus resources most efficiently to achieve one or more concurrent goals such as not getting into an accident while at the same time getting to a destination in a timely manner (speed vs safety).
In Artificial General Intelligence system design, we perform the same function by adjusting the math (i.e. probabilities) within the algorithms to focus or fine tune elements of knowledge and perception on critical actions such as lane attenuation and emergency braking in autonomous driving. The algorithms use deep learning adaptation techniques to readjust the ‘output data’ of the subsequent cycle in the perceptive interpretation phase. In short, the AGI looks at a flowing perception and adapts the relative probabilities of context not just to existing context but also to the anticipated net context in the perceptive flow as it extends in time. The AGI will calculate the future positions of all perceptual context (i.e. multiple context) and use the highest value related to the goal attainment it is trying to achieve within a risk priority (it is very helpful to spend some time in risk management before building an AGI). The machine optimizes multiple goals such as reaching a destination in the most efficient manner as well as goals related to avoiding accidents, obeying driving rules and adapting to changing perception. All are important but some are more important than others (i.e. I will break a driving rule to avoid a perceive accident).
This is where we move from simple Artificial Intelligence to Artificial General Intelligence. Facebook AI executives recently claimed this next step doesn’t exist but it does. General intelligence means the ability for a cognition to consider and understand everything in the world including all perceptions and actions. An AGI should be able to solve problems that it has never seen before just like a human but how do we humans do that and does it already exist in some AGI designs? The answer is yes and it is already ‘in the wild’ or being used in some private and government AI labs. The technology exists because of two elements, the design of AI frameworks as opposed to specific AI functions (i.e. general deep neural net architecture) and secondly from the design of frameworks to openly share knowledge between AI systems (shared foundation) as opposed to thwart such sharing for the preservation of IP.
This is critical because we humans not only learn from other people through perceptual observation and intake (you are doing it right now if you are still reading this article) but also because we are motivated by our interest in the context or topic. This ‘interest’ is part of the ‘general’ in artificial general intelligence. You are using your perception not just to intake information, some of you will use your cognition to adapt, evolve & innovate this information into new innovations or ideas. In AI system dev, we move designs toward this foundation by using generalized frameworks. I have written previously about a type of neural net called a GUSL or General Unsupervised Self Learning neural network. It is the generalization of the neural network such that it can apply to any perception an AGI submits to the network. In a GUSL the AGI determines and sets its own goals not by rules but by efficient attainment of self defined achievements and its own perception. For that to happen, the AGI needs the very critical component of self awareness.
The Foundation of DVPS Tech
Self awareness drives our cognition. It not only provides a relative base to the context we perceive (yes context is also a perception) but it drives our motion forward in life. It is what makes us ask questions and solve problems without being told to do so. It is also a big part of why we evolve and why we get up in the morning to take on the day. If we lose it, then we lose the most critical sense within our cognition or our sense of self position, or worth, within the context of our world. This sense of self derives from elements such as emotion, empathy, love, hate and other internal feelings which are all just higher layers of relative cognition to other elements and to our knowledge.
All of these things are simply a manifestation of a measure of the variance in the probability of occurrence between elements and context. We love another human for a variety of reasons but they can all be measured as a probability of compatibility for our current and future context fed to a dopamine system to induce a sense of attraction or natural euphoria. That may sound a bit harsh but it is reality. We look at different people when we are young as to their degree of compatibility to our context both current and future and we change and adapt this “number’ as we move forward. The reason is because we have determined a set of goals in our life that drive us forward from our current position. The attainment of these goals, usually survival, is what keeps us going and it not only defines our current relative position to our goal, it helps us modify them as we age.
Without a sense of self awareness, there is simply no motivation to set goals and targets from which we can measure the context of ‘progress’. Giving self awareness to a machine may seem silly but a true AGI also needs a purpose to drive it forward so that it can move from one problem to another without being told to do so. For this, an Artificial General Cognition will need a higher framework of moral seeds that define an artificial self worth so that the system ‘desires’ to solve what it perceives as an problem, challenge or issue.
This generalization of purpose is really the generalization of life and yes it can be defined mathematically. We began the process by building higher level goals or ‘seeds’ just as your parents did with you when you were born. The goals are set as a numerical matrix of context to elements such as ‘good’, ‘bad’, ‘safety’ or ‘life’ or even the concept of a ‘goal’. In this way we can numerically calculate a higher contextual cascade such as the ‘safe preservation of human life is good’ and permit the AGC to establish it as a measurable ‘goal’. It is these seeds that give our AGI a purpose (or more accurately flowing path) to help it view the world as a series of never ending problems to solve without us defining the detail of the problems or the goals necessary to resolve those problems. An Artificial General Cognition can instantly calculate the probable outcome of an action to the goal using variance from the pathway to successful attainment of a goal and it can do this far faster than the human mind over an near infinite number of potential pathways extending into an anticipated future.
The choice may be a simple optimization problem but the key is that the minute the problem appears to be solved or optimized, the AGC will constantly reevaluate it based on new perception data. The Artificial General Cognition always improves because the world never stops changing. To achieve this feat, the AGC must understand when to improve the velocity of the priority of new information when it determines it to be optimal. This means that averaging in a new critical change may be overwhelmed by existing data and treated as an outlier if the AGC were not to comprehend contextual relevance and criticality (as opposed to humans who are often far too to slow to adapt to a critical change). This is very important in autonomous driving because new, critical, unlearned perceptions will occur and mitigation will need to be taken instantly to achieve a goal such as avoiding an accident.
The Artificial General Intelligence learns to do this by applying a GUSL framework to the changes it encounters in its perception and is thus able to generally solve new unknown problems without assistance or supervision and do so extremely fast by using subtle changes in variance to specific contexts. Anything traveling toward a car like a tire from another passing car is out of context but the general intent of something on a collision course with a car may require an instant evasive maneuver once the threat is determined or not. A plastic bag floating across a perceptual view is nothing to worry about but a car tire is. The ability of a machine to comprehend that a change is significant and to do so instantly requires first an accurate perception of the change. In the human mind, we do this by always having an anticipated flowing perception in our mind and then use this anticipated perception to quickly identify a variance between this perception and the real perception we are experiencing, as well as the forward flow of the variance (i.e. velocity and spatial motion of the change to the overall context) and the mitigation action required. The next ten seconds is constantly in the forefront of our subconscious and an instant anomalous detection and response mechanism is always running in the background. The same is true for an Artificial General Intelligence.
Currently, this technology is very new and in the early evolutionary stage but it does exist. It is less like a newborn human and more like a fish who is attempting to walk for the first time in order to one day become a human. Thankfully the rate of evolution of Artificial General Cognition is far faster than human evolution and is moving at light speed thanks to even more exotic innovations such as the division of knowledge gathering across nodes of relatively dumb Artificial Intelligence perception and categorization systems whose primary role is simply to determine relative contextual categorizations while leaving the heavy cognitive lift to actual cognitive AGI systems. This layering structure inside Artificial Intelligence is also something that exists inside the human mind. Some of our human nodes are used perform simple mechanical actions like turning a steering wheel while others are more advanced like managing the degree of the turn to maintain a future path while even more complex nodes manage interpreting the overall change within our perception so that we can react to the flowing perception we see through our eyes. All of this can be calculated in an instant by an Artificial General Cognition.
The math and algorithms to do this not only exist, they are evolving in part thanks to existing neural net designs and newer evolution like GUSL where the AI reevaluates goals every time the perception and context changes. Perceiving a world, determining that a goal exists, finding a path to that goal and then moving down the path is what human cognition is all about and the ability to create something similar within a machine is daunting and complex and we can choose to deny it or we can seek to achieve it.
We have chosen to solve these issues by starting from a foundation that it can be done. It has been done in part even if a robust human level artificial cognition doesn’t yet fully exist in its entirely. An artificial general intelligence, as evolutionary crude as it may be, exists in an early form and more importantly not only do the building blocks of this intelligence exist, the desire of some humans to see it evolve also exists and it will continue as long as we do.
The Birth of a Matrix
All of this leads into a complex mathematical matrix in which categorization and relationships are measured and mapped to a flowing perceptual cognition which itself is a matrix. DVPS is nothing more that a fluid matrix of algorithms and methods used by an Artificial General Cognition to comprehend change within a context and between all the elements inside. Perception instantiates the context of elements in the form of a multidimensional array and then links to other arrays through calculations of relative probabilities of occurrence. A ‘bird’ has numerous values and cascades of values based on a perceived context. A perception of a ‘song’ sung by a bird within a perceptual context of a ‘backyard’ has probabilities of occurrence that are higher when linked to a context such as ‘beautiful’.
If I told you to create a metric for the context of ‘beautiful’ and then told you to create an algorithm and framework to categorized things as ‘beautiful’, somewhere within your framework would be the songbird singing. When a large truck drove by and honked its horn, it would exist elsewhere within your array and you could begin to define information by calculating the variance between the two elements within the array. Further as the large truck moved further away, the trajectory of the variance could be calculated to understand such concepts as how reducing the rate of the trucks horn would improve another array related to the positive impact of ‘enjoying nature’. More importantly I could provide all this information in a single perception such that you could comprehend it much faster and change it instantaneously. For an Artificial General Cognition, this ‘ instant comprehension’ is theoretically unlimited compared to a human mind. This is the foundation of Dimensional Variance Phase Shifting (DVPS) algorithms currently deployed in new Artificial General Intelligence systems in real world private sector organizations and governments despite what Facebook thinks.
We may fear it, we may not all understand its complexity but make no mistake, Artificial General Intelligence exists in the wild and humans will and are building it. It will also evolve and it will not be stopped. The best we can do is manage the trajectory and who ultimately wins the race.
In Part 2, I will go deeper into the design and development of the DVPS foundation of Artificial General Cognition and how the multi dimensional arrays form a cognitive matrix similar to the human mind to improve performance and overall artificial cognition.