Reliability
Out of Distribution
Explore model behavior outside the shape of training data.
Training Support
Blue points are observed data; green is the true function; red is the network prediction.
Neural Network Extrapolation
Watch the model fit data in-distribution and fail more sharply outside it.
Function
Noise level
Training MSE
0.012
Out-of-distribution MSE
1.520
The network sees training data only inside x in [-1, 1]. Its in-distribution error falls as training progresses, while OOD error remains high because the observed context is narrow.
Implementation note
OOD behavior is where demos often fail. The model can look smooth inside training support and become arbitrary outside it.