Multi ­scale Stacked Sequential Learning

Place: Large Lecture Room - CVC

Affiliation: Computer Vision Centre and Universitat de Barcelona  

The classification and or segmentation of sequential data (such as text, audio, images and videos) can be improved by moving from i.i.d. classifiers to contextual‐aware methods. One popular approach to contextual-aware classification is the Conditional Random Fields (CRF), or other related graph-based algorithms. In this seminar we will show a method, called “Multi-scale Stacked Sequential Learning” (MSSL) which allows considering the context in a very efficient way. The method proved to outperform CRFs in several tasks, in terms of accuracy, precision, sensitivity. Training and testing times are more than one order smaller than CRFs. We tested the method on text classification, image segmentation (binary and 8 classes problems), temporal classification, and volumetric classification.   Moreover, its implementation is straightforward (less than half an hour in MATLAB and less than 100 lines of code) and, being a meta-learning scheme, any classifier can be employed in its use (kNN, Adaboost, SVMs, etc…). The meta-learner MATLAB code will be provided to the audience.