Just Pick a Sign: Optimizing Deep Multitask Models with Gradient Sign Dropout

TL;DR

Proposes Gradient Sign Dropout (GradDrop), a probabilistic masking method to resolve gradient conflicts, improving multitask and transfer learning performance.

cs.LG 🔴 Advanced 2020-10-14 25 views
Zhao Chen Jiquan Ngiam Yanping Huang Thang Luong Henrik Kretzschmar Yuning Chai Dragomir Anguelov
multitask learning gradient optimization transfer learning deep learning algorithm improvement

Key Findings

Methodology

GradDrop calculates gradient sign consistency probability P = 0.5 * (1 + Σ∇Li / Σ|∇Li|) and masks conflicting gradients probabilistically. It ensures consistent updates while retaining inter-task competition.

Key Results

  • On CelebA, GradDrop reduced error rate from 8.71% to 8.52% and improved F1 score from 29.35 to 29.57, outperforming PCGrad and GradNorm.
  • In CIFAR-100 transfer learning, GradDrop combined with GradNorm achieved 28.9% Top-1 error and 1.01 test loss, outperforming baselines.
  • On the Waymo Open Dataset, GradDrop improved 3D APH from 53% to 56%, significantly outperforming other methods.

Significance

GradDrop addresses gradient conflicts in multitask learning, providing more stable convergence and significant performance improvements. Its modular design allows seamless integration into existing frameworks, benefiting both research and industry.

Technical Contribution

GradDrop introduces gradient sign consistency probability P to probabilistically mask conflicting gradients, avoiding performance drops caused by gradient cancellation. It guarantees joint loss minima and demonstrates broad applicability across multitask and complex single-task models.

Novelty

GradDrop is the first method to use gradient sign consistency probability for conflict masking, retaining inter-task competition unlike PCGrad, while offering theoretical guarantees.

Limitations

  • Performance may degrade in high-noise or sparse gradient scenarios due to slower convergence.
  • Requires additional hyperparameter tuning, such as leak parameter `li`, to balance task priorities.
  • Higher computational demands, especially on large datasets.

Future Work

Future research could explore GradDrop's applications in self-supervised and reinforcement learning, as well as optimizing its computational efficiency and automating hyperparameter tuning.

AI Executive Summary

Multitask learning often suffers from gradient conflicts, leading to unstable training. Existing methods like PCGrad resolve conflicts via projection but suppress inter-task competition, limiting performance.

This paper introduces Gradient Sign Dropout (GradDrop), which calculates gradient sign consistency probability P to probabilistically mask conflicting gradients. GradDrop is a modular layer compatible with any deep network, incurring no additional inference cost. Theoretical analysis shows GradDrop ensures stable convergence to joint loss minima.

Experiments demonstrate GradDrop's superiority on CelebA, CIFAR-100 transfer learning, and the Waymo Open Dataset, especially in complex 3D object detection tasks. Despite computational demands, GradDrop offers a novel optimization perspective with broad application potential.

Deep Analysis

Background

Multitask learning optimizes multiple tasks by sharing network parameters, but gradient conflicts often hinder convergence. Existing methods like GradNorm and PCGrad address conflicts by adjusting gradient magnitudes or directions but face performance bottlenecks.

Core Problem

Gradient conflicts in multitask learning cause parameter updates to cancel out, degrading performance. Effectively resolving conflicts while retaining inter-task competition is a key challenge.

Innovation

GradDrop introduces gradient sign consistency probability P to evaluate gradient alignment. By probabilistically masking conflicting gradients, it avoids 'gradient tug-of-war' and ensures theoretical joint loss minima.

Methodology

  • �� Compute gradient sign consistency P: P = 0.5 * (1 + Σ∇Li / Σ|∇Li|).
  • �� Generate mask M based on P: retain gradients with high consistency, mask opposing directions.
  • �� Apply M to gradient updates, ensuring consistent update directions.
  • �� Optional leak parameter `li` allows partial gradient flow to prioritize specific tasks.

Experiments

Experiments on CelebA, CIFAR-100, and Waymo Open Dataset validate GradDrop's performance in multitask, transfer learning, and complex single-task models. Baselines include PCGrad, GradNorm, and MGDA.

Results

On CelebA, GradDrop achieved 8.52% error and 29.57 F1. In CIFAR-100 transfer learning, it achieved 28.9% Top-1 error. On Waymo, it improved 3D APH to 56%.

Applications

GradDrop is applicable to multitask learning, transfer learning, and complex single-task scenarios like 3D object detection, image classification, and NLP tasks.

Limitations & Outlook

GradDrop may struggle in noisy datasets or sparse gradient scenarios. It also requires hyperparameter tuning and has higher computational costs.

Plain Language Accessible to non-experts

Imagine a kitchen with multiple chefs sharing one pot. Each chef has their own recipe, but if they all add ingredients randomly, the dish might taste terrible. GradDrop acts like a smart assistant, deciding which chef's ingredients to add based on how well they align with the overall recipe, ensuring a delicious final dish.

ELI14 Explained like you're 14

Think of playing Minecraft with friends, where everyone wants to build their own house but shares the same materials. If everyone grabs materials randomly, no one finishes their house! GradDrop is like a manager who helps share the materials fairly so everyone can build their dream house. Cool, right?

Glossary

Gradient Sign Dropout

A probabilistic masking method that resolves gradient conflicts by evaluating and masking inconsistent gradient directions.

Used to improve multitask learning performance.

P (Gradient Sign Consistency Probability)

A metric ranging from 0 to 1 that measures the consistency of gradient directions.

Determines which gradients to mask.

PCGrad

A method that resolves gradient conflicts by projecting gradients to avoid interference.

Used as a baseline for comparison with GradDrop.

CelebA

A dataset with 40 binary attributes derived from celebrity facial images.

Used to evaluate GradDrop in multitask learning.

Waymo Open Dataset

A large-scale 3D point cloud dataset for autonomous driving research.

Used to test GradDrop in complex single-task models.

Open Questions Unanswered questions from this research

  • 1 How can GradDrop's computational efficiency be improved for large-scale datasets?
  • 2 What is GradDrop's performance in self-supervised or reinforcement learning scenarios?
  • 3 Are there alternative masking strategies that outperform gradient sign consistency?

Applications

Immediate Applications

Multitask Learning Optimization

GradDrop can directly improve multitask models in image classification, object detection, and more.

Transfer Learning

GradDrop effectively resolves conflicts between source and target tasks in transfer learning.

Long-term Vision

General Deep Learning Optimization Module

GradDrop could become a standard module in deep learning frameworks, applicable across tasks.

Abstract

The vast majority of deep models use multiple gradient signals, typically corresponding to a sum of multiple loss terms, to update a shared set of trainable weights. However, these multiple updates can impede optimal training by pulling the model in conflicting directions. We present Gradient Sign Dropout (GradDrop), a probabilistic masking procedure which samples gradients at an activation layer based on their level of consistency. GradDrop is implemented as a simple deep layer that can be used in any deep net and synergizes with other gradient balancing approaches. We show that GradDrop outperforms the state-of-the-art multiloss methods within traditional multitask and transfer learning settings, and we discuss how GradDrop reveals links between optimal multiloss training and gradient stochasticity.

cs.LG cs.CV