What is basic intelligence (harmonics of knowledge)?

In the AI world, we believe intelligence to be answering questions such as, “given the current state of the scene around me should I accelerate?” in a self-driving program, or in a medical analysis program “given the person’s test results, does the person have diabetes?” or “Is the payment transaction on this credit card fraudulent?” or “Will this user be interested in this ad for a washing machine?” and many many such types of so-called intelligent questions. While finding the answer to these types of questions can be one of the outcomes expected from a “knowledge and intelligent system”, calling the answering of such questions itself as “intelligence” is wrong. I find this is the reason we end up with the kind of rigid AI we have because we have tightly added learning logic that can search for answers to pre-determined questions. To compensate for the rigidity introduced, we start talking about AGI or general intelligence.

To draw a parallel, while the functional outcome of the various components within a food processor is to cut vegetables or knead flour, the core of the operation is in the circuit that controls the speed of the motor that rotates the cutting blades. Change the cutting blades to various forms, and we achieve multiple actions. In my view, decision making, predictions and all of the other things that are focussed on by our current AI learning algorithms are such higher-level functions that should be got by assembling together one of more basic intelligence. We should not focus on writing algorithms that answer these types of questions but ask ourselves what basic intelligence is and how that can be translated to answer these questions, just as how the motor rotation has been translated to cutting and dicing vegetables.

In my previous blog, I had said, algorithm need not always be a pre-programmed sequence of instruction set that is executed by a processor. Algorithms can also be represented by a chain of triggers that drive changes in the current structure or other molecular structure or properties. When the triggers of the chain of changes are strung together, it becomes an algorithm representative of a logic. Another example of such a logic can be seen when water drips regularly on a stone or cement floor or any such surface that yields to forces. Based on how the various molecules on the surface react, the slope of the surface, the weak spots on the surface, the forces acting on the surface and water molecules, a path gets established over a period of time and etched on the surface. Subsequently, unless there is a change in the parameters of that system, the water flows in that etched path. The path formed is the intelligence by an algorithm established using the parameters that affect the water flow. Such is the logic that exists all through nature and in the analysis we do. If we can emulate this intelligence then can we get the natural AI that does not seem to skim the surface of intelligence.

To create this type of algorithm, we need to start at the basics: data. As I have said in many previous blogs we can use the accumulative nature of things around us to create “observers” rather than writing complex algorithms as in a logic-based system. That accumulative observer automatically creates a representation for “knowledge” present in the data we want to process. This encodes various attributes as variations in different properties of the output of the “observer” that can be something like molecular structure. The continuity is represented as the number of bonds formed, the values represented as strength of the bonds and the various parameters observed together as the structure. The data is now in the state of encoded knowledge rather than raw data, which may not seem much. It is just bringing the various parameters values and the duration of their change together into “knowledge”. Anything we do with this knowledge becomes intelligence because we are not collecting the information directly from data.

To understand the next step, we need to understand the nature of the knowledge formed using this process. In the current system, given, the data we collect using IoT sensors or any other means are scalar in nature, we tend to stamp the data with “time” and use this as a driving factor to relate data. When we do this, we make a very huge assumption: that, just because parameters were measured at the same time, they are all a result of the same cause. This need not necessarily be true. To compensate we need to add external intelligence into the system to prevent relating irrelevant data. Hence rigidity starts creeping in. The knowledge representations that we create using the natural accumulation of structures have encoded in them the duration over which the accumulation occurs. Each individual knowledge representation is built in parallel with each other reacting to various relevant parameters as and when they occur, automatically encoding the occurrence time. Thus, the primary nature of the encoded knowledge is that the it has become “time irrelevant” for further processing.

To understand “why irrelevancy of time is important”, let us take another common application of AI, namely, medical prognosis, example, prognosis of heart attack. In the current systems, the medical information present belong to patients who have already had a mild heart attack. This data does not have the information related to the trigger and accumulation over the trigger due to various causes that lead to the tipping point that causes heart attack. Using purely this data to detect whether another person can possibly have a heart attack or not is what we call guesswork. What we need is knowledge that accumulates along with the gradual degeneration of the various parameters of the body that can be analysed. For example, it is possible that there was a gradual degradation of the heart valves that was triggered by a single or collection of stress and accumulated as the stress levels grew for the person. Here, not all degrading parameters found in the body at some instant t1, need to be related or be indicative of a heart attack. In fact, we would always find that some parameter that changed at some previous time ‘t1 – X years’ is more related to the state of the current parameter at t1, rather than another parameter at the same time t1.

So, data relations that need to be established is not between parameters measured purely at the instant time ‘t1’, but has to be across “knowledge” collected, related by chaining of events. While this is one example, this is true of all the knowledge and intelligence that we as humans have. We can relate seemingly irrelevant information to form a sequence and derive conclusions from it. This “chaining of knowledge” can be based on many distribution parameters. For an image, it is chaining of spatial distribution, for medical history, it is “chaining of events”, for language processing, it is “chaining of grammatical sense”. This is the process of “sequencing of knowledge” and is the first basic intelligence that needs to be built for an AI.

Taking the example, medical prognosis of heart attack further, chaining the events until the current collected knowledge gives us just the information at the current state that can indicate anything. To give any prognosis, we need to subsequently continuously monitor related information and map the behaviour of the information before we can give a prognosis. Similar is the case with analysing the image. Say we wanted to use the “processing of the scene” around us to drive a car, then we need to take the information gathered from the chaining of spatial distribution and continuously monitor and map the behaviour of the scene to take a decision of how to react. Creating a “continuous map of the behaviour of the sequence of knowledge” is the second basic intelligence that needs to be built for an AI.

This process of “sequencing and knowing the behaviour of knowledge” can be seen similar to the “harmonics that we get in music (the scale, the tempo and so on)”. If there is no harmony in the notes used, the mistake of the music seems glaring to a lot of us and becomes noise. In the classical carnatic music it is called the Arohanam (ascending scale) and Avarohana (descending scale) of notes that needs to be in a raga. This is the first and basic intelligence that needs to be built for building any intelligent system. Create and know the harmonics of the knowledge collected. The next step to this should be drawing conclusions from this behaviour and thus answering the questions.