Failure as a Fingerprint of Complex Systems


Failure replication

Two systems can succeed at the same task yet succeed in completely different ways: planes are good for flying, but their flight is very different from bird flight. Conversely, two systems that fail differently at the same task are unlikely to be working in the same way, as is now recognized by Artificial Intelligence (AI) researchers. Yet, the focus of AI is still overwhelmingly on reproducing some tasks seen as hallmarks of human success: writing and other language-adjacent abilities, image recognition, image production. When AI fails at these tasks, it fails in spectacularly different ways than humans do.

Both engineering and natural sciences have embraced purposeful failure as a tool to deepen their understanding of complex systems. In neuroscience and psychology, one of the most
compelling ways to understand how the brain works is to study how it fails. Brain damage, irrational decisions, sensory illusions: internal or external changes that make the brain fail are how we find how the brain succeeds. Failure is used to understand complex systems beyond neuroscience: reverse-engineering computer software, understanding animal behavior, identifying solid materials... Failure even defines Science itself. An hypothesis is considered scientific if and only if it is “falsifiable”: if it can reproducibly fail.

Both engineering and natural sciences have embraced purposeful failure as a tool to deepen their understanding of complex systems. In neuroscience and psychology, one of the most compelling ways to understand how the brain works is to study how it fails. Brain damage, irrational decisions, sensory illusions: internal or external changes that make the brain fail are how we find how the brain succeeds. Failure is used to understand complex systems beyond neuroscience: reverse-engineering computer software, understanding animal behavior, identifying solid materials... Failure even defines Science itself. An hypothesis is considered scientific if and only if it is “falsifiable”: if it can reproducibly fail.

Recent work on visual illusions, including my own, provides examples of this approach. Illusions are a fascinating example of sensory failure: they reveal conflicts between the information volunteered by independent processes in our brains. Other animal species can also be fooled by visual illusions, suggesting a fundamental cause.
The fact that an artificial system based on neural networks can also be fooled by a certain type of visual illusion, and even create its own illusions that fool humans (Sinapayen (2021)), is a good example of shared failure point between biological and artificial systems. Despite the black-box nature of these networks, our research tells us that interactions between our visual environment and the image prediction abilities of our brains could be the root cause of visual illusions.

上へ戻る