We address the problem of extending a pretrained large language model to a new domain that was not seen at training time, like adding a language for which the original model has seen no or little training data. Popular solutions like fine-tuning or low-rank adaptation are successful at domain adaptation, but formally they do not add any extra capacity and degrade the performance in the original domain. Our paper analyzes this extension problem under three angles: data, architecture and training procedure, which are advantageously considered jointly. In particular, we improve adapters and make it possible to learn an entire new language while ensuring that the output of the neural network is almost unchanged in the original domain. For this purpose, we modify the new residual blocks in a way that leads each new residual block to output near-zeros in the original domain. This solution of neutral residues, which borrows architectural components from mixture of experts, is effective: with only 20% extra learnable weights compared to an original model trained on English, we get results that are significantly better than concurrent approaches (fine-tuning, low-rank or vanilla adapters) in terms of the trade-off between learning a new language and not forgetting English.
View on arXiv