With multilingual machine translation (MMT) models continuing to grow in size and number of supported languages, it is natural to reuse and upgrade existing models to save computation as data becomes available in more languages. However, adding new languages requires updating the vocabulary, which complicates the reuse of embeddings. The question of how to reuse existing models while also making architectural changes to provide capacity for both old and new languages has also not been closely studied. In this work, we introduce three techniques that help speed up effective learning of the new languages and alleviate catastrophic forgetting despite vocabulary and architecture mismatches. Our results show that by (1) carefully initializing the network, (2) applying learning rate scaling, and (3) performing data up-sampling, it is possible to exceed the performance of a same-sized baseline model with 30% computation and recover the performance of a larger model trained from scratch with over 50% reduction in computation. Furthermore, our analysis reveals that the introduced techniques help learn the new directions more effectively and alleviate catastrophic forgetting at the same time. We hope our work will guide research into more efficient approaches to growing languages for these MMT models and ultimately maximize the reuse of existing models.

Multilingual machine translation suffers from negative interference across languages. A common solution is to relax parameter sharing with language-specific modules like adapters. However, adapters of related languages are unable to transfer information, and their total number of parameters becomes prohibitively expensive as the number of languages grows. In this work, we overcome these drawbacks using hyper-adapters -- hyper-networks that generate adapters from language and layer embeddings. While past work had poor results when scaling hyper-networks, we propose a rescaling fix that significantly improves convergence and enables training larger hyper-networks. We find that hyper-adapters are more parameter efficient than regular adapters, reaching the same performance with up to 12 times less parameters. When using the same number of parameters and FLOPS, our approach consistently outperforms regular adapters. Also, hyper-adapters converge faster than alternative approaches and scale better than regular dense networks. Our analysis shows that hyper-adapters learn to encode language relatedness, enabling positive transfer across languages.

Knowledge Transfer has been applied in solving a wide variety of problems. For example, knowledge can be transferred between tasks (e.g., learning to handle novel situations by leveraging prior knowledge) or between agents (e.g., learning from others without direct experience). Without loss of generality, we relate knowledge transfer to KL-divergence minimization, i.e., matching the (belief) distributions of learners and teachers. The equivalence gives us a new perspective in understanding variants of the KL-divergence by looking at how learners structure their interaction with teachers in order to acquire knowledge. In this paper, we provide an in-depth analysis of KL-divergence minimization in Forward and Backward orders, which shows that learners are reinforced via on-policy learning in Backward. In contrast, learners are supervised in Forward. Moreover, our analysis is gradient-based, so it can be generalized to arbitrary tasks and help to decide which order to minimize given the property of the task. By replacing Forward with Backward in Knowledge Distillation, we observed +0.7-1.1 BLEU gains on the WMT'17 De-En and IWSLT'15 Th-En machine translation tasks.