Core Concepts
Emergence of causality from association-oriented training is influenced by data heterogeneity, algorithm stochasticity, and model over-parameterization.
Stats
"Large language models trained with regression loss can unveil causal associations."
"Running large-batch Stochastic Gradient Descent can successfully drive the solution towards the invariant, causal solution."
Quotes
"Causality emerges from association-oriented training due to data heterogeneity, algorithm stochasticity, and model over-parameterization."
"Implicit bias arises from the symbiosis between data heterogeneity and modern algorithms."