Google

Strategy A: Replace 15% of the words in the document, chosen at random, each with a single [MASK] token.
Strategy B: Replace three separate, contiguous spans of words (which together make up 15% of the document&#x27;s total words) with a single [SPAN] token for each span.

In denoising autoencoding, corrupted text can be represented using placeholder slots for the masked or deleted tokens. The model is then trained to fill these slots with the original tokens by leveraging the surrounding context. A key benefit of this method is that using placeholders can result in shorter input sequences, which improves the computational efficiency of the training process.

Training Efficiency in Denoising Autoencoding

Based on the following scenario, propose a specific change to the team's input corruption strategy that would likely increase computational efficiency. Explain the principle that makes your proposed change effective.

Optimizing Training Efficiency

When using a denoising autoencoding objective, a common technique to improve computational efficiency is to replace a contiguous span of several words in the input text with a single placeholder token. Explain the trade-off involved in choosing the length of the text spans to replace. Specifically, what is the benefit of using longer spans, and what is a potential negative consequence for the model's learning if the spans become too long?

Efficiency vs. Learning Trade-off in Denoising

A machine learning engineer is training a model to reconstruct a document from a corrupted version. They are considering two different strategies for creating the corrupted input:

*   **Strategy A:** Replace 15% of the words in the document, chosen at random, each with a single `[MASK]` token.
*   **Strategy B:** Replace three separate, contiguous spans of words (which together make up 15% of the document's total words) with a single `[SPAN]` token for each span.

Assuming all other factors are

Learn Before

Related