Only Birds can Fly
My name is Aamir Mirza, a Data Scientist working at MYOB Australia. In this article my focus will be on the achievements made in related field and how advances made in Artificial Intelligence (AI) are challenging our most deeply held beliefs.
To fully unpack the core ideas behind Machine Learning and AI let’s step back in time to a point when humans first walked on the face of the earth and saw birds fly, it must have been astonishing sight to see such elegant creatures soaring high above them, however over many thousands of years it reinforced the belief that only birds can fly. This belief was so strong and profound that it kept humans from fully exploring man made flight for thousands of years. Humans traditionally have been to a large extends been constrained by our own assumptions and beliefs on how things should be and have shown resistance towards new exploration.
AI are not constrained by such beliefs and are showing us that when it comes to what is possible we haven’t even scratched the surface yet. This century the most profound idea regarding new discovery are no longer coming from humans but AI. Today’s AI is nascent, and it will be sometime before its potential can be fully realized, however even in these early days it is showing us the way forward in ways which are nothing short of astonishing. To explore the core contributions of AI and its impact I have devised this blog in self-contained sub articles which highlight specific achievements and related fields.
The brilliance of Move 37
In 2016 Deep mind create an AI to play the game of GO, the AI initially learned the basic concepts of the game from games found on the internet, once the was sufficient knowledge build AI then further evolved through self-play as in playing the game with a copy of itself. Although this topic deserves a blog in itself my focus would be on the core game play discoveries which came our way as result of such endeavor. The AI completely rejected the notion of traditional game play and started making moves which according to professional GO players looked weird. Looking from a human point of view outside observers doubted whether such an AI will truly be able to compete against top professional players.
In the second match with Lee Sedol, came the brilliant Move 37 by GO playing AI (Alpha GO). The move was so unusual and child like that commentators were dumbfounded on what to say next. However, when Lee Sedol later joined the game it took him a few minutes to truly understand and appreciate the absolute brilliance on Move 37. It negated the entire game play and allowed Alpha Go to win the game in ways humans would have never considered. Upon further introspection, the move was so unusual that one in 10,000 games player would have ever considered that game play, however to the AI — why not?
The Go playing AI forced Lee SeDol to redefine what it means to play Go with an AI Agent unbounded by human presumptions, to win the next games in ways the AI would have not considered. The Series was won 4 to 1 in favor of the AI.
The AI finally showed us that even after thousands of years of GO play, we have only explored a small fraction of what is truly possible. For those of you who wish to engage in some visual delight here is the link to the original move 37 game play.
The curious case of the growing pot
While AI playing games like Chess or Go are a major achievement, however we still have not fully explored what AI is capable of. So let up the ante a bit and talk about games which truly require human level intuition, matter of chance and bluff.
Welcome to Poker playing AI. For those of you not familiar with card game poker, it is usually played between two players (Texas no limit hold’em poker), this kind of game playing in AI is called imperfect information game, rational being unlike Go or Chess where the entire game state is fully visible to both player, in poker each player is aware of their own cards and not of the opponents. In the terms of poker, the AI had to learn from scratch when to hold’em and when to fold’em.
So, what can a world class Poker playing AI teach us about own presumptions? Quite a few things. For those world class players which went up again the AI there were some real shocker in store.
For AI to understand poker it had to master 10 to the power 160 decision point, that is 10 with 160 zeros in front of it. It more human terms this number is bigger than atoms in the known universe. In more human terns it like paying 13 billion hands.
Bluff is not a unique human skill, it is a state of an imperfect information games like poker and can be computational explained and mastered just like any other process. In fact, AI was much better at bluff then its human counterparts.
AI found some unique way to grow the pot (total money of the table) at a much faster rate than any human player before. The key to grow the pot was some very unusual bet sizes AI considered. In a traditional games payer will never consider those bet sizes but to the AI tasked with being the best poker player in the world — why not?
As the AI was operating on a game size of 10 to the power 160 decision point just when humans thought that they had the agent figured out the AI came up with entirely new style of poker play. Here I have included some links to the interview with it developers and actual AI game play.
Click here to go to the developers interview.
Watch the actual game play here.
Perfect and Imperfect information games.
Both player have full visibility over the entire board state (there is nothing hidden). In imperfect information games one or more games states in always hidden. Therefore, much harder to both humans and AI to learn.
These are just games
So far, we have only looked at games as possible application so what is AI and machine learning is capable of. This of course is a valid question, for those reading this blog the question of real world applications must have crossed your mind. Hold on to your hats as in this section we shall look at some amazing breakthroughs attribute to AI.
Using Evolutionary Algorithms for Satellite Antenna Design.
As most of us are aware that launching satellite in space is an expensive business, more the weight of the satellite more the cost. With this in mind NASA engineers turn to a branch of AI called Evolutionary Design. This basic premise to computationally design algorithms which mimic nature way of biological evolution, start with some basic population, breed, mutate and cross over those solution and though the process of nature selection come up with an answer.
What NASA was trying to reduce the size and weight of satellite antenna and still be able to maximize the radiation pattern and communication footprint.
Look — we have a weird looking antenna!
This is the antenna AI came up with:
It has the best radiation pattern, fraction of the size of any traditional antenna human could come up with. While as humans are stuck in our own version of only birds can fly
what this Evolutionary AI taught us, is the repeating pattern — why not!
Full Wikipedia article can be found here.
What about things more down to earth?
As much as AI researchers and Data Scientist bask in the glory of their newly create toys how about things closer to home.
In the process of new medical advances and drug discovery, AI is used to screen through 10 to the power 60 possible candidate molecules related to breast cancer. In the world of medicine this is called Eroom’s law and the reverse of Moore’s Law, it is becoming increasingly harder both scientifically and economically to come up with good candidate molecules. AI is used heavily to come up with good candidate solutions much faster than any human could have.
To give you a more specific example, let’s look at the case if IBM Watson AI in the area of Medical diagnosis and treatment. In Japan a woman was suffering from a rare form
of leukemia and its resistance to traditional treatment, when the doctors turned to IBM Watson for possible solutions.
To find out more about the cause of her illness, the researchers supplied the woman’s genetic data, and Watson cross-checked it with the database of over 20 million oncology records, detecting gene mutations that are unique to a leukemia.
This patient had mutations in more than 1,000 genes, but many of them were not related to her disease and they were just hereditary characteristics she had inherited from her parents, while it would have taken a long while for human scientists to check which of the 1,000 genetic changes were diagnostically important or not, Watson did it in 10 minutes.
Based on Watson’s analysis, the team concluded that the case was one of rare secondary leukemia caused by myelodysplastic syndromes, a group of diseases in which the bone marrow makes too few healthy blood cells.
The doctors changed the patient’s therapy plan, after which her condition improved significantly.
An ocean of possibilities
What the AI is repeatedly showing us that we, with all our intelligence have been swimming in a small pond of possibilities and solutions, whilst there is an ocean of knowledge waiting to be discovered.
For the AI no problem is too small to consider, no alleyway is too dark too explore.
What a wonderful time to be a Data Scientist and AI researcher at MYOB. Towards a future where birds are not the only things that can fly.