Machine learning or parameter fitting is said to be poorly constrained if there are many different values of the weights or parameters that fit, or nealy fit the training data. This can occur of there is insufficient data, but may also bd related to the structure of the model's architecture. In particular the inner layers of deep neural networks are often poorly constrained, especially in the early stages of training, in part because they are far from the clamped inputs and outputs and in part because network symmetries may mean that permutations of the inner nodes are effectively identical.
Used on page 151