Transformer-Patcher: One Mistake worth One Neuron (2301.09785v1)

Abstract: Large Transformer-based Pretrained LLMs (PLMs) dominate almost all NLP tasks. Nevertheless, they still make mistakes from time to time. For a model deployed in an industrial environment, fixing these mistakes quickly and robustly is vital to improve user experiences. Previous works formalize such problems as Model Editing (ME) and mostly focus on fixing one mistake. However, the one-mistake-fixing scenario is not an accurate abstraction of the real-world challenge. In the deployment of AI services, there are ever-emerging mistakes, and the same mistake may recur if not corrected in time. Thus a preferable solution is to rectify the mistakes as soon as they appear nonstop. Therefore, we extend the existing ME into Sequential Model Editing (SME) to help develop more practical editing methods. Our study shows that most current ME methods could yield unsatisfying results in this scenario. We then introduce Transformer-Patcher, a novel model editor that can shift the behavior of transformer-based models by simply adding and training a few neurons in the last Feed-Forward Network layer. Experimental results on both classification and generation tasks show that Transformer-Patcher can successively correct up to thousands of errors (Reliability) and generalize to their equivalent inputs (Generality) while retaining the model's accuracy on irrelevant inputs (Locality). Our method outperforms previous fine-tuning and HyperNetwork-based methods and achieves state-of-the-art performance for Sequential Model Editing (SME). The code is available at https://github.com/ZeroYuHuang/Transformer-Patcher.

• The paper presents Transformer-Patcher, a method for sequential model editing that incrementally adds neurons to correct errors in deployed Transformer models.
• It introduces patches in the final FFN layer, enabling adaptive error corrections while preserving performance on non-edited examples.
• Experimental results show that Transformer-Patcher outperforms existing methods in precision, generality, and efficiency across classification and generation tasks.

Transformer-Patcher: A Technique for Sequential Model Editing in Transformer-Based PLMs

The paper introduces a novel method, Transformer-Patcher, designed to enhance the reliability and robustness of Large Transformer-based Pretrained LLMs (PLMs) utilized in NLP tasks. These models, despite their extensive training and widespread success, are not immune to errors when deployed in real-world applications. Addressing this shortcoming, the authors propose Transformer-Patcher, a model editor that allows for sequential adjustment of Transformer models by incrementally adding neurons to handle ongoing mistakes in the deployment phase.

Key Contributions

1. Sequential Model Editing (SME): The authors recognize the need for continual correction in models post-deployment, terming this process Sequential Model Editing (SME). Unlike traditional Model Editing, which typically focuses on isolated error corrections, SME requires fixing errors sequentially as they arise. This paradigm is vital for practical AI deployment, where models need to adaptively correct their behavior over time.
2. Transformer-Patcher Design: Transformer-Patcher addresses the SME task by introducing additional trainable neurons, or "patches," into the final Feed-Forward Network (FFN) layer of the Transformer model. These patches are designed to modify the model's output selectively without altering the original parameters, thereby preserving the model's performance on non-edit examples. This approach ensures low computational cost and efficient error correction.
3. Performance and Evaluation: The paper presents rigorous experimental results demonstrating Transformer-Patcher's effectiveness across classification and generation tasks using BERT and BART models. It achieves a remarkable ability to conduct thousands of corrections (up to 1000 on classification and 2300 on generation tasks) while maintaining the model's generalization capabilities and performance on previously unseen data.
4. Comparison with Baselines: Compared to existing methods such as fine-tuning and HyperNetwork-based approaches, Transformer-Patcher offers superior precision (Reliability), the ability to generalize to equivalent inputs (Generality), and preserves performance on non-equivalent inputs (Locality). It delivers state-of-the-art results in the SME context.

Implications and Future Directions

The research presented in this paper signifies a practical advance in model editing, introducing a scalable and efficient approach for handling errors in large NLP models. Transformer-Patcher's ability to incrementally and non-intrusively adjust model parameters makes it a promising tool for real-time application deployment, potentially extending to various AI domains beyond NLP.

From a theoretical perspective, the paper highlights the importance of dynamic, non-static model architectures that can adapt post-deployment with minimal disruption to pre-existing learned behaviors. This could inspire future work on more modular and adaptive neural network designs capable of learning from ongoing real-world interactions.

In terms of future developments, enhancing the efficiency and scalability of the patching process to handle even larger models like GPT-3 or similar architectures could be an attractive research direction. Additionally, exploring the integration of this patching mechanism in other deep learning models outside the NLP domain could broaden its applicability and utility.

Overall, Transformer-Patcher presents an innovative solution addressing a critical gap in the reliability of deployed AI systems, setting a foundation for further explorations into self-adaptive neural networks.