通过多视图信息瓶颈学习鲁棒表征_learning robust representations via multi-view inf

论文题目：Learning Robust Representations via Multi-View Information Bottleneck

Summary

• 论文通过将两个视图学习得到的公共信息作为有用表征，将两个视图不共享的部分信息看作是冗余信息，最终两个视图之间相互学习得到标签信息丰富和鲁棒性强的表征。

Problem Statement

• 论文旨在通过优化信息瓶颈理论，构建损失函数使得表征能够含有更多的标签信息和鲁棒性。

Method

• 通过理论证明，得到两个损失函数，一个用于学习表征，一个用于去除冗余信息：
  

Evaluation

• 多视图数据集：直接提取多视图信息
• 单视图数据集：通过数据增强，构建两个视图，再在两个视图上进行学习。

Conclusion

• 通过同时增加标签信息和减小冗余信息，可以学习得到较好的表征。

Notes

• 信息平面：
  

References

• Learning Representations by Maximizing Mutual Information Across Views.
• Learning deep representations by mutual information estimation and maximization.
• How do humans sketch objects?
• Contrastive Multiview Coding.
• On Mutual Information Maximization for Representation Learning.
• Multi-view learning overview: Recent progress and new challenges.

Code

• 两个变量之间的互信息估计：

# Auxiliary network for mutual information estimation
class MIEstimator(nn.Module):
    def __init__(self, size1, size2):
        super(MIEstimator, self).__init__()

        # Vanilla MLP
        self.net = nn.Sequential(
            nn.Linear(size1 + size2, 1024),
            nn.ReLU(True),
            nn.Linear(1024, 1024),
            nn.ReLU(True),
            nn.Linear(1024, 1),
        )

    # Gradient for JSD mutual information estimation and EB-based estimation
    def forward(self, x1, x2):
        pos = self.net(torch.cat([x1, x2], 1))  # Positive Samples
        neg = self.net(torch.cat([torch.roll(x1, 1, 0), x2], 1)) #roll
        return -softplus(-pos).mean() - softplus(neg).mean(), pos.mean() - neg.exp().mean() + 1
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

• 多视图信息瓶颈表征学习模型：

from training.multiview_infomax import MVInfoMaxTrainer
from utils.schedulers import ExponentialScheduler
###############
# MIB Trainer #
###############
class MIBTrainer(MVInfoMaxTrainer):
    def __init__(self, beta_start_value=1e-3, beta_end_value=1,
                 beta_n_iterations=100000, beta_start_iteration=50000, **params):
        # The neural networks architectures and initialization procedure is analogous to Multi-View InfoMax
        super(MIBTrainer, self).__init__(**params)

        # Definition of the scheduler to update the value of the regularization coefficient beta over time
        self.beta_scheduler = ExponentialScheduler(start_value=beta_start_value, end_value=beta_end_value,
                                                   n_iterations=beta_n_iterations, start_iteration=beta_start_iteration)

    def _compute_loss(self, data):
        # Read the two views v1 and v2 and ignore the label y
        v1, v2, _ = data

        # Encode a batch of data
        p_z1_given_v1 = self.encoder_v1(v1)
        p_z2_given_v2 = self.encoder_v2(v2)

        # Sample from the posteriors with reparametrization
        z1 = p_z1_given_v1.rsample()
        z2 = p_z2_given_v2.rsample()

        # Mutual information estimation
        mi_gradient, mi_estimation = self.mi_estimator(z1, z2)
        mi_gradient = mi_gradient.mean()
        mi_estimation = mi_estimation.mean()

        # Symmetrized Kullback-Leibler divergence
        kl_1_2 = p_z1_given_v1.log_prob(z1) - p_z2_given_v2.log_prob(z1)
        kl_2_1 = p_z2_given_v2.log_prob(z2) - p_z1_given_v1.log_prob(z2)
        skl = (kl_1_2 + kl_2_1).mean() / 2.

        # Update the value of beta according to the policy
        beta = self.beta_scheduler(self.iterations)

        # Logging the components
        self._add_loss_item('loss/I_z1_z2', mi_estimation.item())
        self._add_loss_item('loss/SKL_z1_z2', skl.item())
        self._add_loss_item('loss/beta', beta)

        # Computing the loss function
        loss = - mi_gradient + beta * skl

        return loss
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49