Google

Because the initial parameters of a pre-trained model—specifically the classifier parameters $$ \omega $$ and the encoder parameters $$ \hat{\theta} $$—are not originally optimized for a specific downstream classification task, the model cannot be applied directly. Instead, a modified version of the model must be created and adapted to achieve accurate results on the new task.

Inadequacy of Pre-trained Model Parameters for Downstream Tasks

A typical approach to adapting a model for a specific downstream task is full fine-tuning, which involves training the model's overall function, denoted as $$ F_{\omega,\hat{\theta}}(\cdot) $$, on a labeled dataset. By treating this adaptation as a common supervised learning problem with explicit labeling, the process updates all initial parameters to produce a set of further optimized parameters, $$ \tilde{\omega} $$ and $$ \tilde{\theta} $$, making the model suitable for the classification task.

Adaptation of Pre-trained Models via Full Fine-Tuning

As an alternative to full fine-tuning, a classifier can be efficiently adapted to work in tandem with a pre-trained encoder by freezing the encoder's parameters, $$ \hat{\theta} $$. This maintains their pre-trained state, allowing the optimization process to focus solely on updating the classifier's parameters, $$ \omega $$, for the specific task.

Freezing Encoder Parameters During Fine-Tuning

A data science team has developed a large language model by training it on a vast corpus of public internet data. The model's sole training objective was to become highly proficient at predicting the next word in a sequence of text. A new project requires a system to classify customer support emails into three categories: 'Urgent Technical Issue', 'Billing Inquiry', and 'General Feedback'. The team lead suggests deploying their existing next-word prediction model directly for this classification task without any modifications, believing its strong general language capabilities will be sufficient. Evaluate the team lead's suggestion. Is this approach likely to succeed? Justify your reasoning based on how the model was originally trained.

Evaluating the Direct Application of a General Language Model

A research lab develops a large-scale language model by training it on a massive corpus of historical literature to become an expert at completing sentences in the style of 18th-century authors. A separate team wants to use this model, without any modifications, to power a modern-day customer service chatbot. Explain the fundamental reason why this approach is likely to fail, focusing on the model's internal configuration.

Mismatch Between Pre-training and Downstream Objectives

A team develops a large language model by training it on a vast collection of text from the internet, with the sole objective of making it proficient at predicting the next word in a sequence. They then attempt to use this model directly, without any changes, to categorize customer support emails into 'Billing Issue', 'Technical Problem', or 'Feature Request'. The model performs poorly. Which of the following statements best explains this outcome?

Learn Before

Related