Controlling Conditional Language Models without Catastrophic Forgetting

By Tomasz Korbak, Hady ElSahar, Germán Kruszewski, Marc Dymetman. 2021

43 citations recorded by the source corpus.

Machine learning is shifting towards general-purpose pretrained generative models, trained in a self-supervised manner on large amounts of data, which can then be applied to solve a large number of tasks. However, due to their generic training methodology, these models often fail to meet some of the downstream requirements (e.g., hallucinations in abstractive summarization or style violations in code generation). This raises the important question of how to adapt pre-trained generative models to meet all requirements without destroying their general capabilities ("catastrophic forgetting"). Recent work has proposed to solve this problem by representing task-specific requirements through energy-based models (EBMs) and approximating these EBMs using distributional policy gradients (DPG). Despite its effectiveness, this approach is however limited to unconditional distributions. In this paper, we extend DPG to conditional tasks by proposing Conditional DPG (CDPG). We evaluate CDPG on four different control objectives across three tasks (translation, summarization and code generation) and two pretrained models (T5 and GPT-Neo). Our results show that fine-tuning using CDPG robustly moves these pretrained models closer towards meeting control objectives and -- in contrast with baseline approaches -- does not result in catastrophic forgetting.

Subfields: Artificial Intelligence, Natural Language Processing, Prompting & Reasoning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

In Prompting & Reasoning

Citations in the corpus

On Reinforcement Learning and Distribution Matching for Fine-Tuning Language Models with no Catastrophic Forgetting (2022, methodology): “Consistent with prior work (Khalifa et al., 2021; Korbak et al., 2022a), we observe that Reinforce is able to quickly achieve high levels of constraint satisfaction, but at the cost of large deviations from a, which translates into significantly decreased diversity of generated samples (in terms of…”
PEER: A Collaborative Language Model (2022, background): “Beyond that, they are hard to control (Korbak et al., 2022) and verifying their outputs is challenging as they often hallucinate content (Maynez et al., 2020; Shuster et al., 2021; Nakano et al., 2021) and lack the ability to explain their intentions.”
RL with KL penalties is better viewed as Bayesian inference (2022, neutral): “Take Generative Distributional Control (GDC) (Khalifa et al., 2021; Korbak et al., 2022a), an approach to ﬁne-tuning LMs that obtains results comparable with KL-regularised RL but minimises a slightly different divergence (forward as opposed to reverse KL).”