#capacity

Question 1: Single choice

A network architecture once reached almost zero training loss on a dataset. Retrained on the same data with a learning rate 1000 times smaller, it ends the same number of epochs with a high training loss. What does the high training loss reflect in this run?

An optimization failure: this run did not reach weights that fit the data1
Too little representational capacity to express a function that fits the data2
Poor generalization from the training examples to new examples3
Training labels so noisy that no model of this size can fit them4

Timed test