MTAN: Multi-Task Attention Network

by Shikun Liu, Edward Johns, & Andrew J. Davison

We present MTAN, a novel multi-task architecture which allows learning of task-specific feature-level attention. Our model can be trained end-to-end and can be built upon any feed-forward neural network, is simple to implement, and is parameter efficient. We show that MTAN is state-of-the-art in multi-task learning compared to existing methods, and is also less sensitive to weighting schemes in the multi-task loss function.

CVPR 2019

Multi-task Attention Network Multi-task Attention Network

Introduction

Compared to standard single-task learning, training multiple tasks whilst successfully learning a shared representation poses two key challenges:

  1. Network Architecture (how to share): A multi-task learning architecture should express both task-shared and task-specific features. In this way, the network is encouraged to learn a generalisable representation (to avoid over-fitting), whilst also providing the ability to learn features tailored to each task (to avoid under-fitting).

  2. Loss Function (how to balance tasks): A multi-task loss function, which weights the relative contributions of each task, should enable learning of all tasks with equal importance, without allowing easier tasks to dominate. Manual tuning of loss weights is tedious, and it is preferable to automatically learn the weights, or design a network which is robust to different weights.

However, most prior MTL approaches focus on only one of these two challenges, whilst maintaining a standard implementation of the other. In this work, we introduce a unified approach which addresses both challenges cohesively, by designing a novel network which (i) enables both task-shared and task-specific features to be learned automatically, and consequently (ii) learns an inherent robustness to the choice of loss weighting.

The Design of MTAN

MTAN consists of two components: a single shared network, and K (number of tasks) task-specific attention networks. The shared network can be designed based on the particular task, whilst each task-specific network consists of a set of attention modules, which link with the shared network. Each attention module applies a soft attention mask to a particular layer of the shared network, to learn task-specific features. As such, the attention masks can be considered as feature selectors from the shared network, which are automatically learned in an end-to-end manner, whilst the shared network learns a compact global feature pool across all tasks.

Produced by OmniGraffle 7.18.3\n2021-03-29 18:16:38 +0000 encoder Layer 1 conv Rectangle merge [1 x 1] conv BatchNorm ReLU [1 x 1] conv BatchNorm Sigmoid Element-wise Multiplication [3 x 3] conv BatchNorm ReLU conv conv pool pool Attention Module for Encoder Produced by OmniGraffle 7.18.3\n2021-03-29 18:16:38 +0000 decoder Layer 1 samp Rectangle merge [1 x 1] conv BatchNorm ReLU [1 x 1] conv BatchNorm Sigmoid Element-wise Multiplication [3 x 3] conv BatchNorm ReLU conv conv conv samp Attention Module for Decoder

Visualisation of MTAN, showing the global / task-shared features (up) and task-specific features (down) from the encoder and decoder part of the network respectively. All attention modules have the same design across all other layers in the network, although their weights are individually learned.

Each task-specific attention module only contains two convolutional layers composed with [1 × 1] kernels, introducing very few parameters for each task. The [3 × 3] convolutional layer represents a shared feature extractor for passing to another attention module, following by a pooling or sampling layer to match the corresponding resolution.

The attention mask, following a sigmoid activation to ensure attended features to be in the range between [0, 1], is learned in a self-supervised fashion with back-propagation. If an attention mask is close to 1, such that becoming an identity map, the attended features are equivalent to global feature maps, and thus the tasks share all the features. Therefore, we expect the performance of MTAN to be no worse than that of the standard multi-task network with hard-parameter sharing, and we show results demonstrating this in the following section.

Experiments

In this section, we evaluate our proposed method on two types of tasks: one-to-many predictions for image-to-image regression tasks; and many-to-many predictions for image classification tasks.

NOTE: We suggest readers to check out this survey, offering the detailed evaluation and analysis of the current up-to-date multi-task architecture and loss function design (including MTAN built on top of a stronger backbone architecture: ResNet-50 with dilated convolution).

Baselines

Most prior multi-task learning architectures are designed based on specific feed-forward neural networks, or implemented on varying network architectures, and thus they are typically not directly comparable based on published results. Our method is general and can be applied to any feed-forward neural network, and so for a fair comparison, we implemented 5 different network architectures (2 single-task + 3 multi-task) based on the same backbone: SegNet, which we consider as baselines.

Dynamic Weight Average (DWA)

For most multi-task learning networks, training multiple tasks is difficult without finding the correct balance between those tasks, and recent approaches (GradNorm, Uncertainty Weighting) have attempted to address this issue. To test our method across a range of weighting schemes, we propose a simple yet effective adaptive weighting method, named Dynamic Weight Average (DWA). Inspired by GradNorm, DWA learns to average task weighting over time by considering the rate of change of loss for each task. But whilst GradNorm requires access to the network’s internal gradients, our DWA proposal only requires the numerical task loss, and therefore its implementation is far simpler.

Update: We recently developed a new and more general optimisation strategy. See Auto-Lambda for more details.

Results on Image-to-Image Predictions

The following table shows experimental results for NYUv2 datasets across all architectures. Results also show the number of network parameters for each architecture relative to single task learning. Our method outperforms all baselines across all learning tasks. Following this, we then show qualitative results on the CityScapes validation dataset. We can see the advantage of our multi-task learning approach over vanilla single-task learning, where the edges of objects are clearly more pronounced.

Type #Params Architecture Segmentation
(Higher Better)
Depth
(Lower Better)
Surface Normal
Angle Distance
(Lower Better)
Within t
(Higher Better)
mIoU Pix Acc Abs Err Rel Err Mean Medium 11.25 22.5 30
Single Task 3 One Task 15.10 51.54 0.7508 0.3266 31.76 25.51 22.12 45.33 57.13
4.56 STAN 15.73 52.89 0.6935 0.2891 32.09 26.32 21.49 44.38 56.51
Multi Task 1.75 Split, Wide 15.89 51.19 0.6494 0.2804 33.69 28.91 18.54 39.91 52.05
2 Split, Deep 13.03 41.47 0.7836 0.3326 38.28 36.55 9.50 27.11 39.63
4.95 Dense 16.06 52.73 0.6488 0.2871 33.58 28.01 20.07 41.50 53.35
3 Cross-Stitch 14.71 50.23 0.6481 0.2871 33.56 28.58 20.08 40.54 51.97
1.77 MTAN (Ours) 17.72 55.32 0.5906 0.2577 31.44 25.37 23.17 45.65 57.48

13-class semantic segmentation, depth estimation, and surface normal prediction results on the NYUv2 validation dataset with equal task wighting. Please refer to the original paper for more results on other weighting schemes.

Input Image

Ground-Truth
(Segmentation)

Single-Task
Learning

MTAN

Ground-Truth
(Depth)

Single-Task
Learning

MTAN

CityScapes validation results on 7-class semantic labelling and depth estimation, trained with equal weighting. The original images are cropped to avoid invalid points for better visualisation. The red boxes are regions of interest, showing the effectiveness of the results provided from our method and single task method.

Robustness to Task Weighting Schemes

MTAN maintains high performance across different loss function weighting schemes, and is more robust to the choice of weighting scheme than other methods, avoiding the need for cumbersome tweaking of loss weights. We illustrate the robustness of our method to the weighting schemes with a comparison to the Cross-Stitch Network, by plotting learning curves with respect to the performance of three learning tasks in NYUv2 dataset.

Cross Stitch Network

Multi-task Attention Network

Equal Weights     Uncertainty Weights     DWA

Validation performance curves on the NYUv2 dataset, across all three tasks (semantics, depth, normals, from left to right), showing robustness to loss function weighting schemes on the Cross-Stitch Network (top) and our Multi-task Attention Network (bottom).

Visualisation of Attended Features

To understand the role of the proposed attention modules, in the following figure we visualise the first layer attention masks learned with our network based on CityScapes dataset. We can see a clear difference in attention masks between the two tasks, with each mask working as a feature selector to mask out uninformative parts of the shared features, and focus on parts which are useful for each task. Notably, the depth masks have a much higher contrast than the semantic masks, suggesting that whilst all shared features are generally useful for the semantic task, the depth task benefits more from extraction of task-specific features.

Input Image

Shared Features

Semantic Mask

Semantic Features

Depth Mask

Depth Features

Visualisation of the first layer of 7-class semantic and depth attention features of our proposed network. The colours for each image are rescaled to fit the data.

Results on Many-to-Many Predictions (Visual Decathlon Challenge)

Finally, we evaluate our approach on the recently introduced Visual Decathlon Challenge, consisting of 10 individual image classification tasks (many-to-many predictions). Evaluation on this challenge reports per-task accuracies, and assigns a cumulative score with a maximum value of 10,000 (1,000 per task).

The following table shows results for the online test set of the challenge. As consistent with the prior works, we apply MTAN built on Wide Residual Network. The results show that our approach surpasses most of the baselines and is competitive with the current state-of-the-art, without the need for complicated regularisation strategies such as applying DropOut, regrouping datasets by size, or adaptive weight decay for each dataset, as required.

Method #Params ImageNet Aircraft C100 DPed DTD GTSR Flower Oglt SVHN UCF Mean Score
Scratch 10 59.87 57.10 75.73 91.20 37.77 95.55 56.30 88.74 96.63 43.27 70.32 1625
Finetune 10 59.87 60.34 82.12 92.82 55.53 97.53 81.41 87.69 96.55 51.20 76.51 2500
Res. Adapt. 2 59.67 56.68 81.20 93.88 50.85 97.05 66.24 89.62 96.13 47.45 73.88 2118
DAN 2.17 57.74 64.12 80.07 91.30 56.54 98.46 86.05 89.67 96.77 49.38 77.01 2851
Piggyback 1.28 57.69 65.29 79.87 96.99 57.45 97.27 79.09 87.63 97.24 47.48 76.60 2838
Parallel SVD 1.5 60.32 66.04 81.86 94.23 57.82 99.24 85.74 89.25 96.62 52.50 78.36 3398
MTAN (Ours) 1.74 63.90 61.81 81.59 91.63 56.44 98.80 81.04 89.83 96.88 50.63 77.25 2911

Top-1 classification accuracy on the Visual Decathlon Challenge online test set.

Conclusion

In this work, we have presented a new method for multi-task learning, the Multi-Task Attention Network (MTAN). The network architecture consists of a global feature pool, together with task-specific attention modules for each task, which allows for automatic learning of both task-shared and task-specific features in an end-to-end manner. Experiments on the NYUv2 and CityScapes datasets with multiple dense-prediction tasks, and on the Visual Decathlon Challenge with multiple image classification tasks, show that our method outperforms or is competitive with other methods, whilst also showing robustness to the particular task weighting schemes used in the loss function. Due to our method’s ability to share weights through attention masks, our method achieves this state-of-the-art performance whilst also being highly parameter efficient.

Citation

If you found this work is useful in your own research, please considering citing the following.

@inproceedings{liu2019mtan,
    title={End-to-End Multi-task Learning with Attention},
    author={Liu, Shikun and Johns, Edward and Davison, Andrew J},
    booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
    pages={1871--1880},
    year={2019}
}