Google

Objective A: Maximize the sum of log Pr(y|x) over all pairs in the dataset.
Objective B: Maximize the sum of log Pr(sequence) over all pairs, where sequence is the concatenation of x and y .

A training objective for a model can be formulated as maximizing the joint log-likelihood of concatenated input-output sequences. For a dataset `D` of input-output pairs `(x, y)`, the optimal parameters `˜θ` are found by maximizing the sum of the log-probabilities of the combined sequence `seq_{x,y}`. The formula is: $$˜θ = \arg \max_{\theta} \sum_{(x,y)∈D} \log \text{Pr}_{\theta}(\text{seq}_{x,y})$$ This approach is equivalent to maximizing the conditional log-likelihood `log Pr(y|x)` when the input distribution `Pr(x)` is not dependent on the model parameters `θ`.

Training Objective as Joint Log-Likelihood Maximization of Concatenated Sequences

A language model can be trained by maximizing the joint log-likelihood of a concatenated input-output sequence, `log Pr(x, y)`. This is often treated as equivalent to maximizing the conditional log-likelihood, `log Pr(y|x)`. Explain the mathematical reasoning and the specific condition required for these two objectives to be equivalent in terms of finding the optimal model parameters.

Equivalence of Training Objectives

Based on the provided scenario, explain why the chosen training objective is leading to poor performance on the intended task. In your analysis, break down the joint probability into its constituent parts and describe how the relative lengths of the input and output sequences affect the model's learning process.

Evaluating a Language Model's Training Objective

A language model is being trained on a dataset of input-output pairs `(x, y)`. Two different training objectives are proposed:

*   **Objective A:** Maximize the sum of `log Pr(y|x)` over all pairs in the dataset.
*   **Objective B:** Maximize the sum of `log Pr(sequence)` over all pairs, where `sequence` is the concatenation of `x` and `y`.

Under which of the following conditions will optimizing for Objective B be mathematically equivalent to optimizing for Objective A?

Learn Before

Related