One Initialization to Rule them All: Fine-tuning via Explained Variance Adaptation

Foundation models (FMs) are pre-trained on large-scale datasets and then fine-tuned on a downstream task for a specific application. The most successful and most commonly used fine-tuning method is to update the pre-trained weights via a low-rank adaptation (LoRA). LoRA introduces new weight matrices that are usually initialized at random with a uniform rank distribution across model weights. Recent works focus on weight-driven initialization or learning of adaptive ranks during training. Both approaches have only been investigated in isolation, resulting in slow convergence or a uniform rank distribution, in turn leading to sub-optimal performance. We propose to enhance LoRA by initializing the new weights in a data-driven manner by computing singular value decomposition on minibatches of activation vectors. Then, we initialize the LoRA matrices with the obtained right-singular vectors and re-distribute ranks among all weight matrices to explain the maximal amount of variance and continue the standard LoRA fine-tuning procedure. This results in our new method Explained Variance Adaptation (EVA). We apply EVA to a variety of fine-tuning tasks ranging from language generation and understanding to image classification and reinforcement learning. EVA exhibits faster convergence than competitors and attains the highest average score across a multitude of tasks per domain.

Parameter Efficient Diff Pruning for Bias Mitigation

In recent years language models have achieved state of the art performance on a wide variety of natural language processing tasks. As these models are continuously growing in size it becomes increasingly important to explore methods to make them more storage efficient. At the same time their increase cognitive abilities increase the danger that societal bias existing in datasets are implicitly encoded in the model weights. We propose an architecture which deals with these two challenges at the same time using two techniques: DiffPruning and Adverserial Training. The result is a modular architecture which extends the original DiffPurning setup with and additional sparse subnetwork applied as a mask to diminish the effects of a predefined protected attribute at inference time.