# Gradient Surgery for Multi-Task Learning

**Authors:** Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, Chelsea Finn  
**arXiv:** [2001.06782](https://arxiv.org/abs/2001.06782)  
**Date:** 2020-01-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2001.06782v1/

## Key points

- Deep learning and deep reinforcement learning systems have achieved remarkable results in areas like image classification, game playing, and robotic control.
- Data efficiency is a significant challenge in these systems.
- Multi-task learning has emerged as a promising strategy to improve efficiency by leveraging shared structures across tasks.
- However, multi-task learning presents optimization challenges that hinder substantial efficiency gains compared to single-task learning.
- A recent study titled "Gradient Surgery for Multi-Task Learning" identified three conditions causing detrimental gradient interference in multi-task optimization landscapes.
- The proposed "gradient surgery" method mitigates interference between task gradients by projecting conflicting gradients onto the normal plane of another task's gradient.
- Extensive experimentation showed significant improvements in efficiency and performance outcomes using this approach on challenging multi-task supervised and RL problems.
- The technique is model-agnostic and can be integrated with existing multi-task architectures for enhanced performance.
- This research sheds light on the complexities of multi-task learning optimization landscapes and provides a practical solution to overcome detrimental gradient interference, paving the way for more effective use of multi-task learning approaches in deep learning and RL systems.

## Summary

In the realm of deep learning and deep reinforcement learning (RL), systems have showcased remarkable achievements in various domains such as image classification, game playing, and robotic control. Despite these advancements, a significant challenge that persists is data efficiency. To address this issue, multi-task learning has emerged as a promising strategy to leverage shared structures across multiple tasks for more efficient learning processes. However, while multi-task learning holds great potential, it also presents several optimization challenges that hinder the realization of substantial efficiency gains when compared to learning tasks independently. The complexities surrounding why multi-task learning poses greater challenges than single-task learning have not been fully elucidated until now. In a recent study by Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn titled "Gradient Surgery for Multi-Task Learning," a set of three conditions within the multi-task optimization landscape were identified as culprits causing detrimental gradient interference. The researchers developed a novel yet straightforward approach to mitigate this interference between task gradients. Their proposed method involves what they term "gradient surgery," which essentially entails projecting a task's gradient onto the normal plane of another task's gradient if conflicting gradients are detected. Through extensive experimentation on challenging multi-task supervised and RL problems, this approach demonstrated significant improvements in both efficiency and performance outcomes. Importantly, the technique is model-agnostic and can be seamlessly integrated with existing multi-task architectures to further enhance overall performance. This innovative research sheds light on the intricacies of multi-task learning optimization landscapes and offers a practical solution to overcome detrimental gradient interference. By addressing these challenges head-on, the study paves the way for more effective utilization of multi-task learning approaches in deep learning and RL systems across diverse applications.