AI and Machine Learning: Moving from Training to Education
The debate of whether AI will ever achieve capabilities at par or beyond human intelligence is ever ongoing. It certainly has intensified with the recent advancements in AI, Machine Learning (ML), and Deep Learning (DL) with some believing that the current technologies are already capable of paving the way for Artificial General Intelligence (AGI). You can hear and read about such debates every now and then which can sometimes create awkward situations even among the best known and reputed researchers in the field.
While motivated by the dream of AI, applications of ML for the most part have focused not on achieving AGI but rather on problems that humans themselves struggle at. For example, looking for patterns that were just either hard to observe by the human eye or difficult to translate into tractable mathematics, models, and codes by the human brain. It is perhaps these needs that drove development of frameworks that not as much enable transference of human intelligence into machines but rather compliment it.
Recent developments of past 10–15 years that enabled availability of exascale amounts of data, plus the hardware that could efficiently store and process such data, paved the way for the Deep Learning revolution that could not have nearly as good an impact 20 years ago. But then came in the publicity and with that followed the hype. Suddenly almost everyone, including those who couldn’t tell the difference between AI and ML/DL, was jumping the bandwagon. Some were looking for solutions to their problems without really understanding what tools will best serve their purpose, others were selling themselves or their solutions without incorporating any insights specific to the problem at hand.
In the midst of the hype and marketing, people (most in numbers, least in expertise) got the idea that perhaps AI has reached a point where you can feed ML/DL algorithms a whole lot of data and wola!, you will have your solution at the other end of the black box. The algorithm will find and generate all the insights, identify patterns, differentiate between correlation and causation, determine real-world connections and implications, and present the best underlying diagnosis or prediction.
Being from an area of statistical analysis myself, I could tell that a significant set of ML algorithms are statistical mathematical techniques that have existed for quite some time, albeit with variations in flavor and nomenclature depending on application and specialization areas. This indeed is opinion that has been echoed by many experts and practitioners in the field. In my research, focus is on using human intelligence to identify useful insights and translate those understandings of the underlying phenomenon into computer models, ‘guiding’ the algorithms on how reliable and useful information could be found or predicted.
In contrast, ML/DL techniques work with three different strategies, all hoping that the algorithm will be able develop some level of understanding on its own. With supervised learning, you are essentially only telling the algorithm what is what out in the real world with labeled training set. With reinforcement learning, you try to correct its behavior when it gets something wrong. The third area of unsupervised learning, with the promise of machines able to train themselves without intervention, is where the progress is most limited. These three strategies are probably not going to be enough for us to come anywhere close to AGI.
Without going into much deeper topic of what intelligence is or the fact that we humans don’t even understand our own intelligence, let alone replicate it in machines; let us consider a very basic example that by relevance should be easy to prove a point.
Suppose you have been blessed with your first child. We all know babies have immense abilities to observe and learn. After a first few months, your baby will start paying attention and you can instruct (or supervise) what it can put in its mouth and what it can’t. That is a good strategy for a few months. During the first few years, you will let your baby crawl, walk, run around, touch, observe, smell, and taste things with you re-inforcing positive and good behaviors and correcting the bad ones. But that’s just not enough after a while. The world is too big to explore and life is too complex to figure out with minimal help so you finally have to think about school!
With school comes education, a process although long and tiresome but irreplaceable at least until the child develops maturity. To ensure the best outcomes, you send your child to a great school because you want nothing less than the best for the development of your child’s mind, knowledge, and skills. But what if the school wasn’t very good? What if it has the same teacher with a high-school diploma teaching maths, history, arts, literature, sciences, programming, etc. Do you think this kind of school would help your child learn in-depth about the world he/she would live in or the one that has a separate teacher with specialization in respective subject areas?
That’s essentially what (majority of) people are trying to do. First, using the same ML/DL algorithms without context and alterations made to tailor a specific application and expecting magical results at the end. Second, the algorithms and frameworks developed for the most part have the approach of training, not education. If this not how we make children (who are undoubtedly way much smarter than even their adult counterparts) learn and excel, then how can one use this approach with computers and expect them to develop any kind of intelligence whatsoever?
Computers and machines are in their very nascent stage of learning and achieving any level of intelligence. It is going to take a variety of techniques to not just train but educate them, much in the same way we educate our young. There are no shortcuts to this process and none are needed. Anything less motivated by quick sales and profits will only lead to sub-optimal performances, risks, and disappointments. Those who realize this are already investing billions of dollars in research which is bound to pay long term dividends.
No matter which side of AI debate you are on, we all should be able to agree that we have yet to achieve a lot. Current developments have certainly been significant but the celebrations should not lead to slowing down of further efforts in developing even better (and perhaps completely new) frameworks that can make the next paradigm shifts possible.
