Companies often find this You optimize modelsAn effective approach to creating a big language model (LLM) fit for purpose and data-driven is to strip the model of a few of its capabilities. After fine-tuning, some models “forget” easy methods to perform certain or previously learned tasks.
Research from the University of Illinois Urbana-Champaign suggests a brand new method for retraining models that avoids “catastrophic forgetting,” through which the model loses a few of its prior knowledge. The article focuses on two specific LLMs that generate answers from images: LLaVA and Qwen 2.5-VL.
The approach encourages corporations to retrain only narrow parts of an LLM to avoid retraining your complete model and significantly increasing computational costs. The team claims that catastrophic forgetting isn’t true memory loss, but relatively a side effect of bias drift.
“Training a brand new LMM can cost thousands and thousands of dollars, take weeks, and emit tons of of tons of CO2. Therefore, finding ways to update existing models more efficiently and effectively is an urgent concern,” the team wrote in Paper. “Using this result, we explore optimization recipes that preserve learning while limiting performance drift.”
The researchers focused on a multilayer perceptron (MLP), the interior decision component of the model.
Catastrophic forgetting
The researchers first desired to test the existence and explanation for catastrophic forgetting in models.
To do that, they created a set of goal tasks for the models to finish. The models were then refined and evaluated to find out whether or not they resulted in significant forgetting. But as the method progressed, the researchers found that the models regained a few of their capabilities.
“We also found a surprising result, namely that model performance in sustained benchmarks dropped significantly after training on the counting task and mostly recovered on PathVQA, one other specialized task that isn’t well represented within the benchmarks,” they said. “While we were conducting the forgetting reduction experiments, we also tried to optimize only the self-attention projection (SA Proj) or MLP layers individually, motivated by the finding that optimizing only the LLM was generally higher than optimizing the complete model. This led to a different very surprising result – that optimizing only the self-attention projection layers resulted in a excellent learning of the goal tasks with out a drop in performance on sustained tasks, even after training all five goal tasks in a single order.”
The researchers said they imagine that “what looks like forgetting or interference after fine-tuning to a narrow goal task is definitely a distortion of the output distribution resulting from the shifting of the duty distribution.”
Tight retraining
This insight turned out to be the important thing to the experiment. The researchers found that optimizing the MLP increases the likelihood of “numerical token issuance and a highly correlated decline in held task accuracy.” It turned out that a model that forgets a few of its knowledge is barely temporary and never long-term.
“To avoid biasing the output distribution, we optimize the MLP up/gating projections while keeping the down projection frozen and find that this achieves similar learning to full MLP tuning with little forgetting,” the researchers said.
This allows for a less complicated and more reproducible approach to fine-tuning a model.
By specializing in a narrow a part of the model relatively than extensive retraining, corporations can reduce computing costs. It also allows for higher control of output drift.
However, research focuses on only two models, namely those coping with vision and language. The researchers noted that they can’t conduct the experiment with other models resulting from limited resources.
However, their findings may be applied to other LLMs, particularly for other modalities.
