Science has proceeded uninterrupted for hundreds of years now, through its progress we have emerged from ignorance and awakened to the reality of our Universe. But scientific advancement is now retarded by a fundamental problem, the scientists. In the near future something as important as science will no longer be left to imperfect and inefficient biological scientists, but will become be the realm of digital scientists. In this fourth installment of Confessions of a Biological Scientist, I will discuss whether computers can really provide the creativity necessary for scientific discovery – can we simulate the eureka moment?
Computers are much better than any human at crunching raw numbers, over the last couple of decades they have also come to excel at storing and recalling massive amounts of information with extremely high fidelity. Where computers still lag behind humans though is their ability to form robust models based on data. Thusfar, computers have proven to be poor pattern recognizers, but this problem is far from intractable and significant progress has been made in recent years.
If you have ever snapped a picture and a box magically appeared around the face of your subject, or used more advanced algorithms that sorts like faces from a batch of photographs uploaded to Google+ or Facebook, then you have had interaction with computer pattern recognition algorithms. Using artificial learning algorithms, computers are taking their first baby steps to recognizing patterns in a way that will allow them true recognition of the objects that they are seeing.
Computers might soon be able to differentiate a cup from a cat, but could they really leapfrog all the way to scientific discovery in the near future? Can computers really express the deep sort of creativity necessary to make scientific progress?
Well firstly they don’t need to, and secondly yes they can (maybe).
Computers don’t need express human-like creativity to make significant scientific progress. Scientists often try to bill themselves as exceptional geniuses, bent over messy desks covered with pages of data and obscure scientific reports, coming to realize unexpected connections and innovative hypotheses. While this view is not an entirely false representation, science can be done in the exact opposite manner (and often is).
As opposed to the rapidfire testing of exciting hypotheses, science can also be the carefully controlled collection of data following minimal perturbation of a complex system. This type of science requires no reall creativity, and would require pattern recognition algorithms no more complex than those that find faces in your facebook photos. When you change input A, what happens to output B through Z. This type of predictable, slow and step-wise science is already quite mechanical and is particularly susceptible to automation.
The prototypical example of this type of science is the work of Dr Ross D King, who developed the robotic scientist known as ADAM. With this robot, Dr. King automated the entire scientific process of developing and testing hypotheses for the role of yeast genes in metabolic pathways (summarized here). This work provides a proof-of-concept that boring, step-wise, science can already be done in an entirely automated manner.
While I describe the kind of science done by ADAM as boring, the implication of it are anything but. The fact is that the majority of science needs to be boring, careful testing of predictable hypotheses in order to disentangle the rules underpinning complex biological systems. Indeed, it might be only through the patience and consistency of a fully-automated scientist that we can understand how the complex interactions of the tens of thousands of proteins present in a given cell.
So if we were to hypothetically unleash robot scientists to perform this kind of simple science without any deep creative abilities, how much progress could we really make – How far does boring get us? In my view, boring science can get us (just about) everywhere we want to go.
To illustrate just how great the progress is that could be made, let me give an example of how a cancer diagnosis could go in an age of robotic science:
You come into the lab complaining of non-specific symptoms. A cross-check with your daily biodata, and advanced diagnostic testing shows a clear sign of early lung cancer. You are put on a standard cocktail of drugs to slow the progression of your disease while vital information is collected. With the aid of a robotically controlled surgical robot, a sample of your cancer cells are then quickly biopsied via nano-surgery and sent to a computerized analysis and treatment lab (CAT-Lab).
Within this lab your cells are analyzed directly for protein and DNA content and within days, CAT-Lab will provide information to begin improving your treatment.
Your cells will then expanded within the lab.Techniques are applied to maintain your cells as close to their original state as possible. These expanded cancer cells are then embedded into a standard matrix of lung cells, to simulate the in vivo environment of your lung. The robotic scientist then tests a huge array of cancer drug cocktails. Through stepwise experimentation, a customized cocktail is developed to maximize cancer cell killing, while minimizing damage to your healthy cells. You are immediately put on a customized combination of drugs that quickly and cleanly eliminates the cancer from your lungs.
While at this point your cancer has more than likely been completely taken care of, CAT-Lab has also performed an analysis of the likelihood of recurrence in your case. The analysis suggests that it would be economical to develop a customized “vaccine” rather than run a small risk of the cancer reoccurring. A clear signature of surface protein expression for your cancer cells has been identified. Targeting this combination of markers, a nano-bot swarm which will instantly kill any cell bearing this signature is developed and placed as a rice sized nodule in your shoulder. Should your cancer reoccur, these nanobots will rapidly eliminate the cancer cells before they can cause disease.
So you can see, scaling current generation scientific techniques to automation level could really lead us to miraculous ends. While the scenario I describe above might sound like science fiction, it is absolutely possible in an age of automated science and does not require any special science to be invented beyond what can already be done today. While the extent of science that will be done might be greater or less than what I have described depending on the efficiency and efficacy of such robotic laboratories, I absolutely foresee this type of analysis to begin in the next decade.
The side-benefit of this kind of approach to customized and automated medical analysis would be that it also provides reams of data for “publication”. Taking that data and figuring out what kind of major discoveries lurk inside of these huge data-sets requires more creativity than would provided by robotic scientists like ADAM, or the CAT-Lab. Boring science can get us to the future, but it won’t ever revolutionize the way we understand the world. So can computers ever recapitulate the creativity that is necessary for scientific leaps exemplified by the great theories like evolutionary theory or relativity?
Can we create a robot genius?
The answer to this question is much deeper and harder than the question of whether robots can do basic science. In my view, creativity is a natural extension of our every-day pattern recognition abilities combined with a bit of randomnes and taken to the logical extreme. What we recognize a genius is simply the recognition of a highly unexpected pattern in the world around us. When we can extrapolate some new information across an opaque gap between old and new knowledge, we call it genius.
So if you can accept my view that our pattern recognition abilities naturally lead to creativity, then it seems obvious that exponentially progressing computers which can already recognize faces will soon match and eventually surpass us on the creative front as well. So can a computer do science? The answer is unequivocally, yes, and this will be enough to get us to the future. But, can a computer be creative enough to be the next Einstein? Maybe – we will just have to wait and see for that one.
It’s interesting to see how many people think the AI behind ADAM is simple or trivial. I have yet to come across one person making such claims who has actually studied the algorithms and the mathematics behind it. I invite you to do so to verify your claims. Do you find quantum mechanics and general relativity simple too? 😉
I don’t think that the algorithms and AI behind ADAM are either trivial or simple. I simply want to point out that the kind of science experiments that ADAM does are relatively boring, repetitive and step-wise. This kind of science does not thrill scientists, or inspire undergrads to get into science, but is nonetheless the most important kind of science.
ADAM is an amazing feat, and I would do not think you can overstate its importance as a milestone in science.