Deep learning with noisy supervision

Publication Type:
Issue Date:
Full metadata record
Central to many state-of-the-art classification systems via deep learning is sufficient accurate annotations for training. This is almost the bottleneck of all machine learning algorithms deployed with deep neural networks. The dilemma behind such a phenomenon is essentially the trade-off between the low expensive model design and the low expensive sample collection. For practical purposes to alleviate this issue, learning with noisy supervision is a critical solution in the Big Data era, since the noisily annotated data on the social websites and Amazon Mechanical Turk platforms can be easily acquired. Therefore, in this dissertation, we explore to solve the fundamental problems when training deep neural networks with noisy supervision. Our first work is to introduce the low expensive noise structure information to overcome the decoupling bias issue existed learning with noise transition. We study the noise effect via a variable whose structure is implicitly aligned by the provided structure knowledge. Specifically, a Bayesian lower bound is deduced as the objective and it naturally degenerates to previous transition models in the case that there is no structure information available. Furthermore, a generative adversarial implementation is given to stably inject the structure information when training deep neural networks. The experimental results show the consistently improvement in the different simulated noises and the real-world scenario. Our second work targets to substitute the previous ill-posed stochastic approximation to the noise transition with a rigorous stochastic reallocation regarding the confusion matrix. This work discovers the reason that causes the unstable issue in modeling the noise effect by a neural Softmax layer and introduces a Latent Class-Conditional Noise model to overcome it. In addition, a computational effective dynamic label regression method is deduced for optimization, which stochastic trains the deep neural network and safeguards the noise transition estimation. The proposed method achieves the state-of-the-art results on two toy datasets and two large real-world datasets. The last work aims to alleviate the difficulty that the ideal assumption on the accurate noise transition is usually not fulfilled and the noise could still pollute the classifier in the back-propagation. We specially introduce a quality embedding factor to apportion the reasoning in the backpropagation, yielding a quality-augmented class-conditional noise model. On the network implementation, we elaborately design a contrastive-additive layer to infer the latent variable and deduce a stochastic optimization via reparameterization tricks. The results on a noisy web dataset and a noisy crowdsourcing dataset confirm the superiority of our model in the accuracy and interpretability.
Please use this identifier to cite or link to this item: