Learning More Universal Representations for Transfer-Learning

By Youssef Tamaazousti, Hervé Le Borgne, Céline Hudelot, Mohamed El Amine Seddik, Mohamed Tamaazousti. 2017

60 citations recorded by the source corpus.

A representation is supposed universal if it encodes any element of the visual world (e.g., objects, scenes) in any configuration (e.g., scale, context). While not expecting pure universal representations, the goal in the literature is to improve the universality level, starting from a representation with a certain level. To improve that universality level, one can diversify the source-task, but it requires many additive annotated data that is costly in terms of manual work and possible expertise. We formalize such a diversification process then propose two methods to improve the universality of CNN representations that limit the need for additive annotated data. The first relies on human categorization knowledge and the second on re-training using fine-tuning. We propose a new aggregating metric to evaluate the universality in a transfer-learning scheme, that addresses more aspects than previous works. Based on it, we show the interest of our methods on 10 target-problems, relating to classification on a variety of visual domains.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

ScaIL: Classifier Weights Scaling for Class Incremental Learning (2020, methodology): “Inspired by works such as [25, 29], we introduce a global score computed with Equation 3.”
A Comprehensive Study of Class Incremental Learning Algorithms for Visual Tasks (2020, background): “currence of new data and their integration in the incremental models. 5. Plasticity (P) - capacity to deal with new classes that are signiﬁcantly different from the ones that were learned in the past [69]. E. Belouadah et al.: Preprint submitted to Elsevier Page 1 of 18 arXiv:2011.01844v4 [cs.LG] 15”
DeeSIL: Deep-Shallow Incremental Learning (2018, background): “re while our best congurations with DFE based on 100 and 1000 classes lose only 22 and 12 points respectively. The gap could probably be further reduced if the feature extractors were more universal [10,11]. This could, for instance, be achieved if DeeSIL’s initial training would be done with an eve”