Will Scientific Research be able to avoid Artificial Intelligence
pitfalls?
It’s now obvious that AI, Machine Learning and Deep Learning are no longer buzzwords as they’re getting more and more present in every industry. Notwithstanding the trend has been overhyped in 2017, we are now certain that these technologies will be ubiquitous by 2020.
Scientific research has not been left behind and AI has been the trigger of a significant change in this domain. Machine Learning techniques such as medical image processing, biological features identification by segmentation and AI-based medical diagnosis are officially adopted in by health researchers with an uncritical approach.
This disruptive effect caused by AI is finally very simple to understand. Machine Learning models follow an unavoidable golden rule: the more (unbiased) data we can feed them, the more accurate and reliable they become. For researchers, the convergence of high-performance computing and Big Data seems to be the Holy Grail and the only possible way for new discoveries. A recent example reinforces this idea with the discovery of thousands of black holes in the center of our galaxy — not discovered by a telescope but only by analyzing long-archived data collected nearly 20 years ago.
This attractiveness is not even a trend. Stepping back in front of these technologies would be considered as a scientific obscurantism. We entered an era where Data seem to have the answers to everything. It’s totally rational that humans — especially researchers and scientists — tend to give more credence to the most experimented agent among others. Let’s suppose a situation in which you have to see a doctor for a health problem that may be serious and you have the choice between consulting an eminent specialist having 20 years’ experience or a young doctor starting out in the profession. Whom would you see? No doubt we would all choose the first proposal and we would be right.
However, Dr. Genevera Allen from Rice University in Houston lately alerted the scientific community by raising issues regarding reproducibility and flawed patterns. Based on the definition given by Steven N. Goodman, Daniele Fanelli, and John P. A. Ioannidis:
“Methods reproducibility is meant to capture the original meaning of reproducibility, that is, the ability to implement, as exactly as possible, the experimental and computational procedures, with the same data and tools, to obtain the same results.”
Reproducibility, as well as repeatability, are compulsory components to validate any critical research findings and neglecting these steps can seriously compromise any hypothesis or conclusion drawn from scientific experience. This problem might be partly explained by the huge amount of data analyzed in which the probability to detect meaningless patterns is fairly high. Even though some practices exist to prevent this complication by computing the frequency with which a relevant pattern — matching the real world — should be detected, this indicator, similar to a true positive threshold value, can’t be reliable. In the end, the statistical approach is just not adapted to successfully address the issue. As long as AI agents have neither knowledge nor model of the world, it will not be possible for them to make the difference between a meaningless and a relevant pattern.
To go one step further, It would be interesting to mention the most abstract field of scientific research which is the theoretical one. Let’s precisely take the example of a sub-area namely Theoretical Physics. This domain of research aims to explore, describe and understand the properties of our world such as physical phenomena on Earth, internal structures of atoms and molecules with Quantum Mechanics and Cosmology which studies the origin and evolution of the universe. Unlike experimental physics, research in Theoretical Physics can only be conducted by using high analytical concepts and methods, and by translating physical phenomena into mathematical models.
In a recent article whose title was “Could Artificial Intelligence solve the problems Einstein couldn’t ?” we can read:
“Names like Planck, Einstein, Heisenberg, Schrodinger, Dirac and more are often hailed as the greatest scientific geniuses of our times as a result. No doubt, they solved some incredibly complex problems, and did so brilliantly. But artificial intelligence, quite possibly, could have done even better.”
We are 2019 and I personally disagree with that opinion. I would reformulate and my question would rather be: “Could Artificial Intelligence also solve the problems that Einstein was able to solve?”
I can only agree with the fact that Machine Learning and Deep Learning are amazingly efficient for many tasks when it comes to analyzing large volumes of data in order to find correlations, probable cause and effect relationships, relevant patterns previously unknown and all that at an unprecedented speed — but we also have to say that we are still very far from Herbert Simon’s prediction during a kickoff conference at Dartmouth College in 1956, announcing that thinking machines able to challenge the human mind were “just around the corner”.
I don’t even aim to mention cases where Deep Learning Neural Networks will tend to fail in a spectacular manner when we don’t satisfy their excessive Data Appetite or when we feed them with biased data. But even best AI still lack essential elements to be considered as really intelligent. Humans can engage in reasoning because they have a common sense. Thanks to that, they can make logical inferences about the world around them. Machines don’t.
A real AI should be able to transfer its learning and its intelligence for solving a problem it hasn’t been trained for. DeepMind, a British artificial intelligence company currently owned by Google highlighted the gap with their experiment with the game Breakout. After being trained, DeepMind’s AI knows how to play correctly but it turns out that if we slightly change the rules by just adding an unbreakable area in the center of the blocks, AI is very slow to adapt these changes. A human player is still better than the machine when it comes to adapting to a new game configuration.
But the main issue I want to highlight is the AI’s inability to generate an assumption stipulating that the expected (most accurate) answer cannot be given because more relevant data to solve the problem might not be present in the training dataset. This is precisely why it inescapably leads to doubts when it comes to imagining an AI having a high level of abstraction. Deep Learning is not synonym of Deep Understanding and there is no Machine Learning able to go beyond the training data we feed them with.
And overall, the biggest dead-end in AI is that an AI doesn’t know that it doesn’t know.
Einstein had no data and above all, there was no equation to solve — He demonstrated that there are no solutions to Maxwell’s equations and from that, start conceptualizing and introducing the Special and General Relativity Theories. As you guess, even if we would provide a large amount of data related to Maxwell’s theory of electromagnetism (which is considered the starting point of Einstein’s reasoning) with petabytes of relevant data to the most advanced AI, it would obviously be impossible to obtain as output a new concept of space and time.
It only remains to hope that the Great Minds of tomorrow will have hindsight as well as the capacity to transcend the existing tools in order to ward off all fear of remaining locked into a system both promising and sterile.
References:
Eileen Meyer, Digging Through Long-Archived Data for the Next Big Discovery in Astronomy (May 15th, 2018), thewire.in
Pallab Ghosh, AAAS: Machine learning ‘causing science crisis’ ( February 16th, 2019), bbc.co.uk
Ethan Siegel, Could Artificial Intelligence Solve The Problems Einstein Couldn’t? (May 2, 2018), forbes.com
Clive Thomson, How to teach Artificial Intelligence common sense (Nov 13, 2018), wired.com
