Seminar

Abstract:

Remaining a hot topic for decades, artificial intelligence encounters a boom as a result of the considerable progress achieved on machine learning. Conventional machine learningrequired specific feature extractor instead of dealing with the row form of natural data.Recently, the big breakthrough of deep learning injects new power to machine learning.Deep learning methods are equipped with multiple processing layers to learn representations of raw data with multiple levels of abstraction. These methods have been successfully applied to many areas such as image classification, gesture recognition, and object segmentation. However, supervised deep learning requires sufficient labeled data,which is expensive and time-consuming. Knowledge-transfer based methods appear to tackle this issue accordingly. Transfer learning has been verified as an effective approach in the knowledge transfer area. It transfers knowledge learned from previous tasks to similar ones by mathematically reducing the divergence of different tasks.A major phenomenon in many machine learning and data mining algorithms is that the network training only works for one specific task.

However, there are many similar tasks in real scenarios. If the knowledge learned from previous tasks can be applied to similar ones,the data resource will be saved effectively. In our research, we focus on building up a robust transfer learning algorithms with deep neural networks for knowledge transfer.In this report, we briefly introduce what we have done in this area and the future plan.Firstly, we combine transfer learning with distance metric learning in a unified neural network to achieve cross-age face recognition. Though considerable promising progress has been achieved on face recognition. Cross-age face recognition remains challenging due to its considerable variation of personal appearances. Distance metric learning methods which measure the similarity among each candidate are widely adopted in face recognition area. We integrate distance metric learning with the convolutional neural network to learn ageinvariant features. This algorithm reduces the dependence on considerable labeled data effectively. Experiments are conducted to show the effectiveness of our approach.

Then we consider to apply transfer learning methods to the multi-robot learning process and extend the cross-age face recognition to blood-related face recognition in the further research. We jointly learn deep features and reduce domain discrepancy in a unified deep neural network. Sequential experiments will be constructed to demonstrate the performance of our methods.