In order to build machines that are truly intelligent, humans need to teach them more about phenomena such as cause and effect.
That’s according to Judea Pearl.
Judea Pearl has argued for many years that artificial intelligence has not managed to reach its true potential quickly enough because it has been stuck for more than a decade or so, in a long rut.
Why should anyone listen to Judea Pearl?
Because he is considered as a pioneering figure in the field of machine learning and artificial intelligence.
How do we progress then?
Judea’s prescription for true progress is for humans to finally teach computers how to comprehend the “why” question.
There is little doubt about the fact that artificial intelligence (and machine learning) owe quite a lot to figures such as Judea Pearl.
Especially when it comes to how intelligent machines have become.
Judea Pearl is responsible for leading efforts (back in the 80s) that enabled all types of machines to use probabilistic techniques in order to reason.
But that has changed.
Now, many consider Judea Pearl as one of artificial intelligence’s sharpest critics.
Judea Pearl has recently written a new book on the subject matter.
The book is titled “The Book of Why: The New Science of Cause and Effect”.
In the book, Judea makes a new argument about the state of artificial intelligence.
He says that researchers have handicapped the field, for now.
By not having a complete understanding of what artificial intelligence is really about.
Artificial intelligence research has presented humans with a lot of challenges in the last three decades or so.
The prime of them is for humans to program computer machines in such a manner that they can associate a given potential cause to a given set of conditions which are observable.
Judea Pearl made his name by figuring out the solution to the above-mentioned problem.
How did he do that?
He used a scheme that now everyone calls the Bayesian networks.
Without going into too many details, readers should know that Bayesian network actually made it fairly practical for all computer machines to do tasks that we take for granted now.
What kind of tasks are we talking about here?
We’re talking about tasks such coming up with an explanation that the patient probably had malaria if given that the patient had just returned from Africa with a couple f body aches and some fever.
Back in the year 2011, for his life’s work, Judea Pearl won the very prestigious Turing Award.
As far as computer science honors go, there is none higher than the Turing Award.
Judea Pearl, as mentioned before, won it thanks to the work he had done in the field.
Coming back to the part where Judea Pearl is at odds with some of the artificial intelligence research community, Pearl feels that the field of artificial intelligence went astray.
It actually got mired in associations that made heavy use of probabilistic techniques.
Nowadays it isn’t uncommon to find headlines touting all the latest and the greatest breakthrough in the field of neural networks and machine learning.
No one needs more material to read about how computer machines have managed to master significantly ancient games.
Using the same techniques, machines have also learned how to drive cars.
All of this doesn’t overwhelm Judea Pearl though.
Because the way he sees progress in artificial intelligence is different from others.
Pearl sees the artificial intelligence of today with all its state of the art products as merely souped-up versions.
Souped-up versions of what?
Of what computers and machines even a generation old could already do:
Search and find masked regularities in a given huge set of some data.
According to Judea Pearl, one could easily attribute the majority of the achievements (some of which are impressive according to Pearl) in artificial intelligence using techniques such as deep learning to one simple step:
Judea Pearl is now 81.
And in his book “The Book of Why” he has managed to elaborate his vision for how machines which are truly intelligent would reason and think.
In the book, Judea Pearl argues that the key for artificial intelligence researchers it to try and replace the way machines currently reason (which is by association) with, what he calls, casual reasoning.
In practical terms, instead of machines having the mere ability to just correlate malaria and fever, computer machines must build up the capacity to use their reasoning in order to understand that it is malaria that causes fever.
Once researchers have managed to put in place such kind of a causal reasoning framework, then it would become possible for computer machines to gain the ability to ask questions which are counterfactual in nature.
In other words, machines would have the capacity to inquire about casual relationships and how they would alter given a specific kind of related intervention.
This, according to Pearl, is what everyone should consider as the cornerstone of artificial intelligence and scientific thought.
Judea Pearl, in his book, also proposed a formal language that would help researchers to make machines think like he wants machines to think possible.
That formal language comes in the form of an updated and 21st-century edition of what researchers call Bayesian framework.
Bayern framework is the same framework that enables machines to use probabilistic techniques to reason and think.
Judea Pearl is very clear on his expectations in his book.
He has full confidence in that when machines gain the ability of causal reasoning they would become as intelligent as human beings.
This would further allow machines to communicate with humans beings in a more effective manner.
Additionally, as Pearl explains in his book, machines (once equipped with causal reasoning ability) would have little problems in achieving the status of moral entities in the world.
In other words, the machines would become moral entities that will have the capacity for evil alongside free will.
Judea Pearl recently attended a conference held in San Diego.
There he sat down with some reporters from Quanta Magazine.
Later he also held some follow-up interviews via phone.
During those interviews, reporters asked him questions regarding the reason he made the decision of calling his new book “The Book of Why”?
He replied that he meant the title of the book as a summary.
A summary of the important work he has devoted his time to for the past 25 years or so.
According to Pearl, most of his work has been about cause and effect.
And what cause and effect would mean in a human’s life.
The applications of it and how do people in the field come up with the required answers to all the questions which are inherently casual.
According to Judea Pearl, it seemed pretty odd that science has abandoned such questions.
Pearl said that was the reason why decided to write the book and try to make up for science’s neglect.
Reporters also asked him if his statement were slightly dramatic in the sense that science was all about cause and effect.
So why would Judea Pearl say something like that?
Judea Pearl, while replying to the question, said that he could not see this kind of noble aspirations in the scientific equations that researchers were writing at the moment.
Pearl also mentioned that he considered the language used in algebra as symmetric.
In other words, if X gave out some information about Y, then effectively Y gave out some information about X.
Judea’s work is more related to deterministic relationships.
He believes that currently, the community did not have a way to take a simple fact and write it in mathematics.
To take an example, there was no way to say that strong winds caused the barometer to show a low reading.
Currently, maths can only explain the above-mentioned example in the other way round, that is, the barometer reading went down and hence there is an upcoming storm.
Pearl also said that mathematics, currently, simply did not develop, what he called, asymmetric language.
Asymmetric language is the very thing that is required to capture the community’s understanding that just because X manages to cause Y, in no way, absolutely means that Y manages to cause X as well.
Pearl did understand that some of his views would sound like horrible things to express against the field of science.
He also joked that if he told such things to his mother, his mother would probably slap him.
Fortunately, Pearl said, science was much more forgiving than his mother.
To put it in simpler terms, since science currently did not have a calculus for relationships that were asymmetrical, science actually encouraged scientists to go ahead and create one.
Pearl had the opinion that this was exactly where mathematics came into the equation.
He also said that it thrilled him when he saw how a pretty straightforward calculus of causation actually solved those problems that many legendary statisticians of the time deemed as unsolvable and/or ill-defined.
He further added that one could achieve all of with a lot of fun and ease that comes with finding proofs using geometry one learned in high-school.
We have already mentioned the fact that Judea earned his name in the field of Artificial Intelligence about two decades ago.
He actually taught machines how they could use probabilistic techniques in order to reason.
Continuing from that, a reporter also asked him questions about how the artificial intelligence field looked like back in the day.
To that Pearl answered that in the early 80s emerging problems in those times were of a diagnostic and/or predictive nature.
So, maybe one could have a doctor who tried to look at a multiple number of symptoms from a given patient’s record and then wanted to come up with a reasonable estimate (the probability, in other words) whether the patient had malaria or not.
The community wanted to have expert systems and automatic systems that could replace professionals.
These professionals could include anyone and everyone including an explorer who searched for minerals or a doctor who diagnosed patients/diseases or essentially any kind of expert who got paid for his/her expertise.
It was at that time that Judea developed the idea of carrying out such tasks probabilistically.
The unfortunate part about the whole situation was that standard probability calculations had this requirement of exponential space and exponential time.
To combat such requirements, Judea developed a scheme known as Bayesian networks.
This scheme had more transparency and required, instead of exponential time, polynomial time.
Judea also addressed questions where, in his new book, he described himself as somewhat of an apostate in the artificial intelligence community.
He said, he described himself as such in the sense that as the artificial intelligence community developed certain tools which enabled computer machines to reason (think) with uncertainty, he left that arena.
And started to pursue a much more challenging mission.
That mission involved enabling machines to reason with cause & effect.
Pearl mentioned that several of his artificial intelligence colleagues had occupied themselves with the uncertainty part.
He still knew of circles of AI research which continued their work on diagnosis only and did not worry about the problem’s causal aspects.
According to Judea, all that these researchers want to do is to first predict well and then diagnose well.
He also gave a relevant example.
The majority of the machine learning research/work that is going on today involves researchers conducting operations in the diagnostic mode.
In such a mode they try machines to label objects such as “tiger” and/or “cat”.
Researchers currently give little to no time to intervention.
What these researchers concern themselves in is to just recognize the given object and then move ahead to learn how the object would evolve in the future.
Pearl said he felt like an apostate when he first developed robust tools for diagnosis and prediction but already knowing full well that his tools merely touched the tip of what we know as human intelligence today.
According to Pearl, if the scientific community wanted machines to have the ability to reason about introspection (questions such as “what if I had gone to university?”) and interventions (questions like “what would happen if we can all just ban cigarettes?”) then it had to invoke, what he calls, causal models.
Pearl doesn’t believe that associations have the capability to offer enough.
According to him, that is not his personal opinion but a mathematical fact.
Reporters also asked him about how excited the community felt about all the possibilities that artificial intelligence would open up and how that did not excite him.
He responded by saying he looked a lot into what researchers were doing with techniques such as deep learning.
But he could not hide the fact that all those researchers were stuck.
Because they did not want move above the level of associations.
To put it in another term, curve fitting.
Pearl said that his views might sound like sacrilege.
After all, what he was trying to convey was that all the supposedly impressive deep learning achievement amounted to just one activity.
That activity was trying to fit the curve to the provided data.
Looking at it from the mathematical hierarchy’s point of view, it did not matter how much skill researcher used to manipulate the given data.
Or even what the researcher read from the given data when he/she tried to manipulate it.
In the end, his/her work still represented nothing more than just a curve-fitting exercise.
Of course, it was a curve-fitting exercise that was nontrivial and complex.
It did not take much time for reporters to come to the conclusion that all the advances in machine learning did not impress Judea.
Especially given the way he talked about activities such as curve fitting.
To that, Judea replied that was not the case.
He said machine learning did impress him.
Because he did not expect machine learning to solve so many of the current problems via complex activities such as curve fitting.
As it turned out, machine learning and curve fitting can solve a lot of problems.
However, Pearl also said that if someone thought about the future, what would be next?
The next step according to Pearl was to have a scientist (a robot one) to have the ability to plan a given experiment and then attempt to find unique answers to questions in the scientific community which are still pending.
Moreover, humans should also have the ability to conduct a kind of communication with a given machine that is somewhat meaningful.
In Pearl’s view meaningful means that the machine communication should match human intuition.
So if humans keep on depriving the machine of their intuition concerning cause and effect, they will never have a meaningful conversation with machines.
Without cause and effect, robots will never tell a human that it should have done better on some task.
Currently, humans can do this easily.
And this is where humans lose a very significant channel of machine communication.
Humans have to find a way to equip computer machines with a new model of the environment.
Because if a given computer machine does not have access to a comprehensive model of reality then humans should not expect the computer machine to exist in that reality and behave in an intelligible manner.
According to Pearl, the very first step towards his objective will probably take place in 10 years time.
By then humans may have programmed conceptual models (task driven and purposeful) of reality.
Latest posts by Zohair (see all)
- Is it Legal To Use a VPN Service? (Depends on Where You Live) - 13 August 2018 9:57 PM
- AI Security: Robots Are Coming To Make Friends With Soldiers - 12 August 2018 4:48 PM
- The FCC Thinks 25Mbps Is Actually A Fast Internet Speed - 11 August 2018 3:07 PM