Man vs. Machine

Decades ago, in the 1960s, the Oregon Research Institute decided to create a simple algorithm – one that judges the likelihood of an ulcer being malignant by considering just seven equally weighted factors. To build it, the researchers consulted doctors and asked them to judge the probability of cancer in 96 different cases of stomach ulcers, mixing up X-ray slides and sometimes showing them the same ulcer twice. This was the input on which the algorithm was to be based on - expert judgement using a fairly small data set (by today's standards), and then cleaning it for human errors. The model was supposed to be a starting point. Nothing groundbreaking. The results, however, shocked everyone.

The doctors, whose inputs were used to build the algorithm, had often contradicted themselves when looking at the same X-ray. Even though this sounds absurd, human biases and memory shortcomings works in strange ways. The model, on the other hand, beat even the single best doctor. Do remember that the model had just incorporated seven equally weighted factors. A real-life doctor might consider many more, and assign unequal weights depending on the circumstance. Despite such constraints, a back of the envelope algorithm was good enough to outdo expert judgment.

Like us, machines also use historical data, or inductive reasoning, to arrive at a judgement in such cases. However, unlike us, they do not suffer from lapses of memory, or psychological biases. Also, in some cases, they have access to much more data. Even the celebrated Nobel Prize winning chemist, Linus Pauling, committed a basic blunder when he tried to arrive at the structure of a DNA. If he had access to a computer, with the data and constraints known to Pauling being plugged in, the computer would have raised a red flag on what he proposed - a triple helix in which the phosphates were held together by a hydrogen bond. Ironically, Watson and Crick confirmed the error by referring to Pauling's classic "College Chemistry" textbook. Importantly, it is not that a computer with sufficient AI would have helped Pauling solve the DNA problem, because he did not have access to X-ray crystallography that Watson, Crick, Wilkins, and Franklin knew about, which was so crucial to solve the puzzle. However, a computer would have prevented him to make a blunder, and sometimes, we are as good as our biggest blunders.

With the kind of computing power and big data we have now, the algorithms will only keep getting better. The more data points computers observe, the smarter they become. Not surprisingly, in 2017, Stanford researchers developed an algorithm that can diagnose pneumonia better than expert radiologists. Notably, AI has made these impressive strides in the field of medical research, where doctors spend years to build expertise. How will it impact other not so complex domains is a fairly simple conclusion.