Scalable and unsupervised classification : reality or myth ?

21.02.2018 11:15 – 12:15

With the advent of deep network architectures many remarkable results have been reported in classification and hypothesis testing applications and estimation problems such as image restoration, denoising, super-resolution and compression. Most of these architectures benefit from the presence of a huge amount of training data per class.
The recent architectures assume end-to-end training where the entire architecture is optimized according to some loss function. Despite some obvious progress, interesting reported results, user-friendly ML libraries developed in TensorFlow, PyTorch and Keras and huge interest from the IT industry and research communities there are a number of questions that still remain either completely unsolved or weekly understood.
At the same time, partial solutions are advocated by unsupervised systems based on generative models such as GANs and VAEs. However, in the most cases the generative models have serious restrictions on the dimensionality of input and require accurate alignment. Additionally, in the most cases the validation of produced results in based on visual assessment.
Among the main difficulties towards large scale applications of these systems are: (a) complexity of training for large dimensional data, (b) need of labeled data for supervised classification, (c) limited amount of classes and trivial encoding methods of class labels, (d) vulnerability to perturbations in input and training data and lack of rejection options in classification space, (e) high sensitivity to geometrical deviations in the input data wrt training sets and (f) need of retraining of the entire network with the addition of new classes.
Not less important factors are security and privacy of data next to understanding what is learned by these systems and how the learned features are changed with the addition of new classes or training examples.
The last but not the least factor is a lack of solid information-theoretic foundations behind the design of deep architectures providing some estimates on the fundamental achievable limits of the performance in terms of accuracy, needed amount of training data, etc.. In this presentation, we will try to consider the existing deep architectures from common perspectives and advocate several solutions that partially address the above problems.

Lieu

Bâtiment: Ecole de Physique

Grand Auditoire A
24, quai Ernest-Ansermet

entrée libre