Ref: JAEINT21_EX_1091
Training plan offered by researcher Javier Portilla Muelas and entitled: Study and application of uncoupled characteristics to deep learning
Registration deadline until April 12
The recent emergence of the so-called "deep learning" through new models of artificial neural networks (ANNs) has been a technological milestone of the first magnitude, with direct application to tasks such as pattern recognition and image processing, where they already represent the state of the art. . The related scientific and applied mathematics community is actively working to better understand the properties of these RNAs. In particular, and despite the spectacular successes achieved in recent years in difficult problems, the following aspects still concern: (a) RNAs typically have a very high number of parameters; (b) In order to train this large number of parameters, it is necessary to present a lot of tagged data (for example, thousands or even millions of images), something that is usually only available to large ICT corporations; (c) Despite the large amount of data used for training, it is possible in many cases to generate adversarial examples (unwanted behavior dependent on small modifications of the images), which demonstrate a lack of robustness due to overfitting and poor generalization. This indicates the need to advance in the introduction of a priori models that allow reducing the degrees of freedom of the ANNs.
On the other hand, at the Institute of Optics we have recently developed a feature decoupling method [1,2,3], which has shown great promise in regression, classification and synthesis tasks. So far we have used our method without learning the characteristics from the data. However, we think its greatest potential will come from including it in deep learning models.
[1] J. Portilla and E. Martínez-Enríquez, "Nested Normalizations for Decoupling Global Features," 2018 25th IEEE International Conference on Image Processing (ICIP), Athens, 2018, pp. 2112-2116.
[2] E. Martínez-Enríquez and J. Portilla, "Deterministic Feature Decoupling by Surfing Invariance Manifolds," ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Barcelona, Spain, 2020, pp. 6049-6053.
[3] E. Martínez-Enríquez and J. Portilla, "Controlled Feature Adjustment for Image Processing and Synthesis," 2020 IEEE 22nd International Workshop on Multimedia Signal Processing (MMSP), Tampere, Finland, 2020, pp. 1-6
For any questions about this offer please write to: javier.portilla@csic.es
More information
Rest of the institute's JAE offers 
