No Real Intelligence in AI - That’s the Problem
The prevailing discourse about artificial intelligence can be summed by using the following one word - fear. One really has to look no further than the plethora of comments posted online or listen to the AI experts speaking on TEDx to understand that many feel genuine apprehension regarding what our society will look like with computers taking charge. Without a doubt, we are standing on the precipice of a technology that has the potential to radically shape society as we know it. It is understandable to feel trepidation about the new possibilities that will shortly be coming our way. But let us anchor our fears on realistic probabilities rather than imaginary doomsday scenarios played out as intellectual exercise. In other words, let us put into perspective the capabilities and limitations of artificial intelligence.
To begin with, let us clarify that artificial intelligence is used as an umbrella term. Within this term are many diverse applications from computer vision to improving company operations using “big data.” Second, the term itself - artificial intelligence - is mired in inaccuracy and contradiction. No offence to John McCarthy, the man who coined the term, but the one thing that artificial intelligence as a concept and application lacks is that of which it is assumed to possess in the first place. In other words, there is no “intelligence” in artificial intelligence. We can pretend all we like that we are teaching our computers to think and learn, but really, we are not. Let’s be honest, there is no cognitive function or cognitive ability or any reasoning ability in these machines.
So how do we create training sets or in laymen terms, how do we teach computers to “see” to “think”, to identify animate objects such as a cat, a car or a smile? Let’s take one AI application, computer vision, as an example. We create neural networks by feeding the computer millions of annotated images that have been modified in some gruelling process that takes far too long and is far too labour-intensive. (If there is someone out there that knows of a way to improve this process - raise your hand now!) The point of this is to create sufficient entropy so that the computer identifies the necessary patterns for it to identify an object.
The computer learns to identify the images it is fed by modeling similar human behaviour with a calculation called loss function. Simply put, with each mistake made by the computer, it modifies its learning path so that it comes closer to the value of 0. Similar to human behaviour, our actions are guided by the risk and reward system insofar as we modify plan A upon our discovery that it a certain action has failed to get us closer to our end goal, which in this case is the value 0.
Despite computer vision mimicking human cognitive abilities, it mimics only a very small component. Unfortunately, artificial intelligence software and artificial intelligence is nowhere near having the same powerful cross domain ability to learn and plan like human beings do. Why? Because the capability to learn and reason remains only with human beings. No amount of data can train or trick the machine into the extremely complex and murky world of human cognitive understanding or functioning.
More than that, computer vision in particular and artificial intelligence in general is still about pattern recognition and scripted responses. That the machine’s identification of a cat originated not from an organic and innate understanding of the world around it, but from the programmer labeling those millions of images as a cat renders the probability that machines can reach the same level of reasoning and intellect as human beings as negligible. And perhaps, this is the greatest trouble with artificial intelligence. That in a situation where the computer must make a split second decision its action will be limited to a particular set of configurations. So when luxury car manufacturer Mercedes announced this year that their automobiles will be programmed to kill bystanders instead of their own passengers this is where the real danger lies or perhaps its greatest shortfall.
An even more troubling scenario is what happens when weapons become autonomous, that is, identifying targets and making decisions about them. Imagine a drone armed with artificial intelligence software. On the one hand, its speed and precision is unmatched by humans. Autonomous weapons will save lives of soldiers. On the other hand, its decisions will always be bound to a set of pre-programmed variables unable to take in subtleties or nuances that are beyond its finite, programmed scope. In contrast to humans, artificial intelligence software is not flexible enough to handle fluid situations. In other words, there is no room for split second judgement or analysis brought on by last minute surprises.
The clear benefit artificial intelligence has on our world is automating processes and procedures for industry across many sectors. For example, in the medical diagnostics field when the identification of cancerous areas of human tissue is done using computer vision, the results are much quicker and more accurate than compared to a medically trained pathologist. Along the same lines, there is a lack of medical expertise in the developing world. The World Economic Forum says there is a 10x-20x shortage of positions in the developing world. To fix this problem, it would take approximately 300 years to fill the empty positions. Replacing a team that used to take 5-6 years takes only 15 minutes if replaced by a computer. Now imagine if we can help to enhance their efficiency by using these artificial intelligence applications. The list of imaginable applications for artificial intelligence is endless. Every sector both within the public and private sphere will be influenced.
It is our duty and responsibility to react to the changes both in a practical and moral way. We recognize that there are potential risks involved and there is potential for abuse. However, the benefits to humanity for artificial intelligence far outweigh the potential for misuse. The more and more our lives become digital, the greater our responsibility to make our technology more responsive.
We have a golden opportunity to re-imagine how we connect to machines and it is in humanity’s interest to seize the moment.