Underfitting

A model too weak or too undertrained to learn even the patterns already present in the training data.

Underfitting is a basic machine-learning concept, the opposite of overfitting: the model's error is already high on the training set, and it does just as poorly on validation and test data. Common causes include insufficient model capacity (too few parameters, too simple an architecture), too little training, a poorly chosen learning rate, too much regularization, or inputs that are missing information the task actually needs. It's diagnosed by watching the training loss: training loss that won't come down is underfitting, while training loss that's low but validation loss that rises again is overfitting. In robot learning, a typical symptom is a policy that fails even on the exact demonstration scenes it trained on — the right fix is usually a bigger model or longer training, or checking whether the observation and action data (camera images, action normalization) is broken, rather than reaching for anti-overfitting tools like data augmentation.

ExampleTraining a very small multilayer perceptron to imitate demonstrations of bimanual clothes-folding: the training loss plateaus at a high value early on, and the policy still can't grasp the fabric correctly even when replayed on the exact scenes it trained on — that's underfitting.

Related

Overfitting · Regularization · Training / Validation / Test Set · Loss Function · Parameter Count (Model Size) · Convergence

Sources

Google Machine Learning Crash Course: Overfitting
Wikipedia: Overfitting（含 Underfitting 一节） (Chinese)