by Sourjyamoy Barman, Junior Year, India
“Are we good at making guesses?”
Well, let me ask you that question again but in a different way:
“Are we good at making guesses about some particular things?”
The answer: yes, we are.
Let me explain it to you using some instances. We are entirely aware of the fact that our future is uncertain, however there are few things we are really good at guessing, like we can’t get our meal back after eating it, our broken arm isn’t going to heal overnight or the vase we broke, isn’t going to get fixed itself right away. Similarly, A.I (Artificial Intelligence) has got the ability to make accurate guesses using the fundamental concepts of Special Theory of Relativity proposed by Sir Albert Einstein. Researchers of the prestigious Imperial College claim that AI can make better guesses through an algorithm of causes and consequences found in the Special Theory of Relativity. Humans possess brains which have evolved over millions of years, due to which they have excellent reasoning and thinking abilities.
We are competent to make good guesses because we have strong intuitions about the causes and consequences. However, computers are not good at Causal Reasoning. This is because the machine learning models are incompetent to figure out why one event should follow another event. Rather they have been designed to work on a series of algorithms. We all are pretty much aware of the fact that A.I is being used in several fields but the above cited problem limits the utilization of A.I in a better way. For instance: people with type 2 diabetes are often overweight and have shortness of breath. But the shortness of breath is not caused by the diabetes, and treating a patient with insulin will not help with that symptom.
Researchers in this field are constantly trying to figure out how AI can excel the process of Causal Reasoning. They have developed ML models which try to analyze the events frame by frame by spotting the similar events in a series of actions. For instance, show the AI, a car turning a corner, it will eventually create the next few frames itself! Surprisingly, such systems are used in self-driving cars. But the process is much more time consuming and complicated than it seems, as the ML model requires millions of frames to generate the next few frames, over that any of the frames consist of glitchy pixels, then it can create a much complex and bigger problem in the next frames. Thus, to avoid such problems, researchers at Imperial College have used the theory in a creative way.
They allowed it to generate a whole range of frames that were roughly similar to the preceding ones and then pick those which were most likely to come next. The AI can make guesses about the future without having to learn anything about the progression of time. Basically, being inspired by the light cones, they developed an algorithm. Light Cones are related with the cause and effects in Space-time which was first proposed by Einstein in theory of Special Relativity. Based on the fact that the speed of light is constant, it tries to show the expanding limits of a ray of light. Truly Reasonable, but how did they apply the principles of this theory on ML models? They basically used the proposed idea on two data sets: Moving MNSIT which consists of short video clips of handwritten digits moving around on a screen and the KTH Human Action Series, which contains clips of people performing any kind of motion like walking, running or waving their arms.
The scientists used the A.I program to generate frames by similarity and then used the light cone algorithm to draw circular boundaries around the objects which could be causally related to the given frame. This is how the A.I was designed to pick the frames which were most likely to follow another frame. It’s been designed in such a way that it can successfully detect a particular object in the given frame. For instance, if it’s asked to detect a person with green outfits from a number of frames, then it will reject the frames which has person with no outfit or outfits of different color. Researchers are still working on this for better and exact experimental results. They had the cones expand at a fixed rate, but in practice, the rate will vary. The proposed method is even much more complicated and involves concepts of thermodynamics. For now, researchers plan to set the size(diameter) by hand. But. With the help of existing and stored data, A.I could determine the size of cones itself.
Predicting the next events is a crucial part or function of such programs. Similar programs have been used in Self-driving Cars which have Autopilot technology.
Generally, these predictive systems will be more accurate if they focus on Cause and effect rather than figuring out the correlation.
Researchers believe that such programs can be used in medical fields to predict how would a patient react to certain medicine or treatment. It’s probable that A.I can also predict how a drug would affect a certain disease.
So, now we can probably conclude that A.I is not just about programming and all. In this cosmos, each and everything is connected, we’re just supposed to explore them. Causal Reasoning is also important to take actions and decisions in real world.
And not to mention, if we ever ask the question ‘Why?’, then we need to understand the cause and effect of a particular event
hARTS magazine 