Learning from incomplete features by simultaneous training of neural networks and sparse coding
Attention : séminaire reporté à une date ultérieure
Title : Learning from incomplete features by simultaneous training of neural networks and sparse coding
Asbtract : In this talk, after giving a review of available theory and algorithms for sparse coding of signals, I will show how to use the sparse representation model to solve the problem of training deep neural networks on datasets with missing features. Datasets with limited, weak, noisy labels or incomplete features represent an important and still open problem for important machine learning tasks such as recommender systems, medical screening, self-driving cars and many others. Usually, the problem of missing features is addressed by preprocessing a dataset to complete its missing values (data imputation). In our recent paper presented at CVPR2021 [1], a new supervised learning method is developed to train a deep neural network using only a subset of features per sample, while assuming sparse representations of data vectors on an unknown dictionary. Sufficient conditions are identified, such that, if it is possible to train a classifier on incomplete observations so that their reconstructions are well separated, then the same classifier also correctly separates the original (unobserved) data samples. Simulation results on synthetic and well-known datasets are presented validating our theoretical findings and demonstrating the effectiveness of the proposed method compared to traditional data imputation approaches and one state-of-the-art algorithm.
[1] “Learning from Incomplete Features by Simultaneous Training of Neural Networks and Sparse Coding”, CF Caiafa, Z Wang, J Solé-Casals, Q Zhao. LLID Workshop at CVPR 2021 (Conference on Computer Vision and Pattern Recognition), New York, USA, 19-25 June 2021.
More information : https://ccaiafa.wixsite.com/cesar