Imitation learning errors grow quadratically with task scope

Evidence

Because a single mistake could cause the system to see things it had never prepared for before, once it makes a first mistake, all bets are off. The error grows with the square of the size of the task. Running for ten times as long would produce a hundred times as many mistakes. — Brian Christian, The Alignment Problem, 1001

Categorization

• XREF: The ‘cascading errors’ phenomenon connects to distribution shift and covariate shift problems in ML; also parallels ‘compounding errors’ in human skill acquisition. — connects to: Self-imitation can bootstrap a machine to superhuman skill — Brian Christian, The Alignment Problem, 1057
• corroborated: 2 independent passages support this claim — Brian Christian, The Alignment Problem, 1001

Broader Connections

• Self-imitation can bootstrap a machine to superhuman skill #complicates (0.75) — Brian Christian, The Alignment Problem, 1057