Should companies base AI systems on knowledge or data? The answer may surprise you: the best AI systems today use both.
If you’ve come to AI in recent years, you might be surprised to learn that early AI applications were almost exclusively logic-based: models were built from facts and rules, not learned from data. MYCIN is an early such AI system from the 1970s. It identified bacteria that caused severe infections, and recommended antibiotics. MYCIN used about 600 rules, and also gained knowledge by querying a physician. This approach varies dramatically from using data to build a machine learning system, which is the most widespread approach today.
Indeed, the pendulum has swung far away from the use of domain knowledge towards a very data-centric, or even data-only, approach, with a tendency to use as little domain knowledge as possible. This data focus has produced some remarkable successes in many fields like computer vision and natural language processing.
As AI matures, and as we look to apply it more broadly in the enterprise, the best systems combine both approaches. And missing domain knowledge can lead to dangerous inaccuracies. For example, this pneumonia diagnosis and treatment system estimated patient mortality risk, and also recommended whether patients should take antibiotics at home or be treated in the hospital. However, an early deployment produced dangerous recommendations for a particular class of patients with Asthma. As it turned out, Asthma, which can accompany pneumonia, is usually identified early, triggering physicians to send asthmatic patients to intensive care, resulting in low death rates for these patients. For this reason, a historic data set used to train this system showed that these patients had a low mortality risk. So, for this reason, the algorithm recommended that these patients take antibiotics at home. If the domain knowledge of such cases—which is widely known by physicians—had been considered during data curation, then this dangerous oversight could have been avoided.
Knowledge can also be applied after the model is built—during inference—as in this study that aimed to reduce the number of mammograms a radiologist must read. The team developed a machine learning classifier to identify normal mammograms, In this example, human knowledge came into play after upstream processing by the AI.