1

As the title said, validation set that is affecting weights. But, it might not be what you think.

While the training set is affecting weights based on sample with backpropagation like these steps:

  1. Forward (inference) a train sample input to the architecture.
  2. Get inferred value.
  3. Compare the inferred value with the training sample ground truth, returning the loss value.
  4. Backpropagate the loss value as a reference to update weight state.
  5. Repeat to number one for the next train sample.

The validation set is affecting weights from a higher level perspective based on epoch and no gradient descent like these steps:

  1. Save the weights of initial epoch.
  2. If the validation set performance of current epoch is lower than previous epoch, then restore the previous weights.
  3. Otherwise, save the current weights.
  4. Repeat for the next epoch.

So, it's affecting weights over epochs, not over batches or samples. So, is the validation set still considered a validation set?

Muhammad Ikhwan Perwira
  • 800
  • 3
  • 10

1 Answers1

2

The validation set is still considered a validation set because its purpose remains evaluation, not training. As you explained the validation set influences training indirectly by acting as a signal for when to stop or revert to previous weights, rather than directly modifying the weights through backpropagation, and something else needs to decide whether to continue training or halt based on validation performance.

In order to avoid overfitting, when any classification parameter needs to be adjusted, it is necessary to have a validation data set in addition to the training and test data sets... Since this procedure can itself lead to some overfitting to the validation set, the performance of the selected network should be confirmed by measuring its performance on a third independent set of data called a test set... An application of this process is in early stopping, where the candidate models are successive iterations of the same network

Therefore in some sense your intuition is correct, validation set does affect trained model though indirectly, and another independent test set is thus required to confirm its possibly overfit validation result.

cinch
  • 11,000
  • 3
  • 8
  • 17