Computer Vision with Maker Tech by Fabio Manganiello

Author:Fabio Manganiello
Language: eng
Format: epub
ISBN: 9781484268216
Publisher: Apress

4.

The network overfits the points in the training set. In this case, you may want to either experiment with a network with a lower number of units/layers or pick a higher value for the regularization rate λ in order to increase the “inertia” of the network against the “swings” in your dataset.

As we have seen, once we have found a way toward convergence given a training set, we have mainly two parameters that we can tune to adjust the performance of the model: the learning rate α and the regularization rate λ. We have seen that α determines how fast the network learns when presented with new data and λ expresses the “resistance” of the network against change. Sometimes the dataset is split into three instead of two in order to separately adjust these two values:First, we train the model on the training set and make sure that its cost function constantly decreases. The goal of this phase is to find the values of the weights Θ that minimize the cost function and a value of α (or a function α(t) that returns α over the iteration t) that is a reasonable trade-off between speed and robustness (expressed as the tendency of the model to converge regardless of the starting point).

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.