One Shot Learning in AI Innovation

Modern algorithmic design far exceeds the limits of human cognition in many ways. Armed with large data sets, programmers promise that their algorithms can better predict which prisoners are most likely to recidivate2 and where future crimes are likely to occur.3 Software designers further hope to use large data sets to uncover relationships between genes and disease that would take human researchers much longer to identify.4

But modern machine learning still cannot effectively match human cognition in at least one crucial respect: learning from small data sets. Young children, for example, master new concepts with startling rapidity and fluency. “[G]iven 2 or 3 images of an animal you have never seen before, you can usually recognize it reliably later on.”5