Multi-Task Learning as Multi-Objective Optimization

Published in

Neural Information Processing Systems(2018)

External Links:

Generate Graph

TL;DR

This paper proposes an upper bound for the multi-objective loss and shows that it can be optimized efficiently, and proves that optimizing this upper bound yields a Pareto optimal solution under realistic assumptions.

Abstract

In multi-task learning, multiple tasks are solved jointly, sharing inductive bias between them. Multi-task learning is inherently a multi-objective problem because different tasks may conflict, necessitating a trade-off. A common compromise is to optimize a proxy objective that minimizes a weighted linear combination of per-task losses. However, this workaround is only valid when the tasks do not compete, which is rarely the case. In this paper, we explicitly cast multi-task learning as multi-objective optimization, with the overall objective of finding a Pareto optimal solution. To this end, we use algorithms developed in the gradient-based multi-objective optimization literature. These algorithms are not directly applicable to large-scale learning problems since they scale poorly with the dimensionality of the gradients and the number of tasks. We therefore propose an upper bound for the multi-objective loss and show that it can be optimized efficiently. We further prove that optimizing this upper bound yields a Pareto optimal solution under realistic assumptions. We apply our method to a variety of multi-task deep learning problems including digit classification, scene understanding (joint semantic segmentation, instance segmentation, and depth estimation), and multi-label classification. Our method produces higher-performing models than recent multi-task learning formulations or per-task training.

Authors

Ozan Sener

3 papers

V. Koltun

29 papers

References61 items

Taskonomy: Disentangling Task Transfer Learning

GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks

Routing Networks: Adaptive Selection of Non-linear Functions for Multi-Task Learning

Automatic differentiation in PyTorch

Dynamic Routing Between Capsules

Multi-Task Learning as Multi-Objective Optimization

Published in

Neural Information Processing Systems(2018)

External Links:

Generate Graph

TL;DR

Abstract

Authors

Ozan Sener

3 papers

V. Koltun

29 papers

References61 items

Taskonomy: Disentangling Task Transfer Learning

GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks

Routing Networks: Adaptive Selection of Non-linear Functions for Multi-Task Learning

Automatic differentiation in PyTorch

Dynamic Routing Between Capsules

Scene Parsing through ADE20K Dataset

Descent algorithm for nonsmooth stochastic multiobjective optimization

One Model To Learn Them All

An Overview of Multi-Task Learning in Deep Neural Networks

Multi-task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics

Gradient-Based Multiobjective Optimization with Uncertainties

Pyramid Scene Parsing Network

A Joint Many-Task Model: Growing a Neural Network for Multiple NLP Tasks

UberNet: Training a Universal Convolutional Neural Network for Low-, Mid-, and High-Level Vision Using Diverse Datasets and Limited Memory

Pareto Frontier Learning with Expensive Correlated Objectives

Trace Norm Regularised Deep Multi-Task Learning

Integrated perception with recurrent multi-task neural networks

Cross-Stitch Networks for Multi-task Learning

The Cityscapes Dataset for Semantic Urban Scene Understanding

MOON: A Mixed Objective Optimization Network for the Recognition of Facial Attributes

Inverse Reinforcement Learning through Policy Gradient Minimization

Deep Residual Learning for Image Recognition

Multi-task Sequence to Sequence Learning

Predictive Entropy Search for Multi-objective Bayesian Optimization

Rapid adaptation for deep neural networks through multi-task learning

Multi-Task Learning for Multiple Language Translation

Learning Multiple Tasks with Deep Relationship Networks

Representation Learning Using Multi-Task Deep Neural Networks for Semantic Classification and Information Retrieval

Deep Learning Face Attributes in the Wild

Policy gradient approaches for multi-objective sequential decision making

Pareto-Path Multitask Multiple Kernel Learning

A Survey of Multi-Objective Sequential Decision-Making

Revisiting Frank-Wolfe: Projection-Free Sparse Convex Optimization

Multi-task learning in deep neural networks for improved phoneme recognition

Cross-language knowledge transfer using multilingual deep neural network with shared hidden layers

Towards Pareto Descent Directions in Sampling Experts for Multiple Tasks in an On-Line Learning Paradigm

Combining a multi-objective optimization approach with meta-learning for SVM parameter selection

Multiple-gradient descent algorithm (MGDA) for multiobjective optimization

Clustered Multi-Task Learning Via Alternating Structure Optimization

A Convex Formulation for Learning Task Relationships in Multi-Task Learning

ImageNet: A large-scale hierarchical image database

A unified architecture for natural language processing: deep neural networks with multitask learning

Multi-Task Learning for Classification with Dirichlet Process Priors

Content-based image retrieval through a multi-agent meta-learning framework

Multicriteria Optimization (2. ed.)

Task Clustering and Gating for Bayesian Multitask Learning

Stochastic Method for the Solution of Unconstrained Vector Optimization Problems

Steepest descent methods for multicriteria optimization

A Model of Inductive Bias Learning

Nonlinear multiobjective optimization