In my previous piece, “How intelligent Is your AI?[i]”  the premise was that an intelligent AI would ‘learn, understand, and reason.’ That has not been the case with existing AI. Yet they are very productive tools and are helping us create and capture tremendous value. PwC estimates the incremental value-add or TFP (Total Factor Productivity) by AI to the global GDP could be worth US$15.7 trillion by 2030. Massive indeed. If only some intelligence was fuelled the numbers could multiply. Here is a narrative for possible building blocks to get there.

On a cold winter day of November 2019, Imperial War Museum (IWM) London was assessing various AI engines to choose one to help its archives become vibrant. A picture of an exploding atom bomb was ingested into the engines. One of the most renowned Deep Learning-based engines identified it as a mushroom! (Turns out, this picture sample had yet to find its way into their library of encyclopedias.) Others did not get it either. One company got it. But it had no library or encyclopedias. It did have some data set, as all AI engines need to ‘learn’ (ingest data.) So, what could have been different from this company’s approach? Turns out that the engine was able to create its own ‘context’ when it was asked to identify the picture. Simply put, when the question was asked it ‘also’ considered the fact that the picture was sourced from the war museum enabling it to ignore mushroom as an option. Learning is critical to context creation, which in turn is critical to understanding, which in turn is critical to reasoning. All three elements need to be addressed.  To do that, I will refer to the human decision-making process and paint a series of ‘what-if’ scenarios with parallels from another empirical subject that gained prominence about the same time as AI, though not as popular, called Neuroeconomics.  

Neuroeconomics is a specialization in the human decision-making process that converges concepts of Neuroscience (biological sciences of neural systems), Psychology (the study of human behavior), and Economics (the idea of rational choices and decisions). This discipline aims to provide a single general theory of human behavior of decision making. It focuses on the processes that connect sensation and action by revealing the neurobiological mechanisms by which decisions are made and integrating elements of rational aspects from Economics as well as behavioral aspects from Psychology. I will not dwell on emotions, societal framework, or other factors that form part of our decisions. The idea is to limit our thoughts to human brain functions of learning, understanding, and reasoning and draw parallels with AI.

Learning

For beginners, our brain is bombarded with an estimated 11 million bits of information every second via our sensory organs, but the brain can ‘consciously’ process only about 40 bits per second[1]. Therefore, even if we wanted, we would not be able to ingest (learn) a library (forget library of encyclopedias). So, we ingest differently – first the basics, next intermediate, and then advanced. Example: the learning structure that, perhaps, went into writing this article- I first learned the English alphabet with its 26 letters when I was in pre-primary or primary school. Then I learned words, composed sentences and grammar, writing a paragraph, and so on. Similarly, I learned basic sciences and mathematics, which led me to advanced levels as well as an introduction to computer science, which led me to AI. The sciences also led to Neuroeconomics. So, when I present my views today my learning for all the years acts as a context and constructs meaning out of all learning I have had so far. Not just that, I am just about using the learning (or ingested data) that I need for this one.

On the other end, the learning in AI is flawed – you must have heard “AI is not magic, to even begin it requires volumes of data.” Why?  So, what-if instead of creating libraries or libraries of encyclopedias our AI learned (ingested data) in a structured manner with basic bearings to form a context?

Understanding

Human beings have a limit to consciously process data despite the 100 billion neuron cells communicating with each other. A computer does not. Therefore, it may seem obvious that we can feed in any amount of data in an AI engine. That is indeed happening, but it is not helping to make it understand the data. It is just neat tagging and classification of data in a library, where you can extract data from the shelf when you want. AI scientists have created various forms of Artificial Neural Networks (ANNs) to replicate human neural networks but the sum effect of all them is to store data in a certain manner that can be retrieved basis the need. Fair. We need those.

But these neural networks do not communicate with each other like human neural networks, which use neurotransmitters to communicate dynamically. So, what-if we had the structured learning data stored in neural networks that are fluid or dynamic structure rather than a pre-fixed one?

Reasoning

If learning and understanding is achieved, the reasoning would be easier because the computing power is far ahead in the game. However, if we limit our input-output algorithm to enable the neural networks to self-create connectivity we are likely to get better outcomes as we (as humans) may have ignored certain aspects due to our biases, which is beyond our control to steer away from. As one reader noted in his comments in the previous article, the AI can possibly connect the data set in an unknown way to find something that has never existed before.

This means we go beyond the correlation that Machine Learning based system can do and get to causality into data. That would be a leap, not just another innovation to solve huge problems.  So, what if we left the AI to connect the dots rather than doing it through our algorithms?

Perceptual decisions

Finally, even before we begin assessing AI’s capability to ‘learn and understand’, we must consider that data is limited to Natural Language Processing (NLP) of raw data. NLPs act as sensory organs to convert the real-world data into 0s, 1s, and their combinations to allow AI to understand and differentiate every real-world composition. The human core brain uses a diffusion model to process sensory information. The model has three pre-conditions: brain decides in a non-random manner, the decision is goal-oriented, and there is a choice. The model stresses that the choice should be made as soon as the difference between the evidence supporting the losing alternative exceeds a threshold – sort of reward (and risk or fear) system. Advanced research implementation is on to use these insights to help people with disability to get sensory information to and from their brain cells and have robotic limbs do the basic motor functions. But where we could find the most value for AI is to take perceptual decisions, where the decision aims to categorize ambiguous sensory information (in computer sciences terminology it is called noisy data.)

If our AI must be intelligent it needs to be capable of making perceptual decisions from abstract information as well. That is one reason why digitization (even before digitalization) of legacy systems remains a challenge at organizations. They are convinced that their ‘data is the new oil’ but when they get into the act they realize ‘data is the new “crude” oil.’ It still needs to be explored, extracted, cleaned, and refined even before you can begin to understand its value. And human bias could have used automated filters in the process as well. So, what if we focused our energies on making AI sensors (NLPs) to read abstract (the crude) without the need for going through the entire value chain of refining?

If these 4 what-ifs can be answered as a whole, an AI – with intelligence- can be deployed much faster and get higher productivity. In my mind, it can be done, and it is already being done.

 

[1] https://www.britannica.com/science/information-theory/Physiology