-
Members: FreeCIS
IEEE Members: Free
Non-members: FreeLength: 00:53:39
24 Jul 2020
Machine Learning in the wild (Barbara Hammer, Machine Learning Group, Bielefeld University)
Neural networks have revolutionised domains such as computer vision or language processing, and learning technology is included in everyday�s consumer products. Yet, practical problems often render learning surprisingly difficult, since some of the fundamental assumptions of the success of deep learning are violated. As an example, only few data might be available for tasks such as model personalization, hence few shot learning is required. Learning might take place in non-stationary environments such that models face the stability-plasticity dilemma. In such cases,
applicants might be tempted to use models for settings they are not intended for, such that invalid results are unavoidable.
Within the talk, I will address three challenges of machine learning �in the wild�:
(i) How to learn reliable given few examples only?
(ii) How to learn incrementally in non-stationary environments where drift might occur?
(ii) How to enhance machine learning models by an explicit reject option, such that they can abstain from classification if the decision is unclear
I will present robust approaches how to address these problems, which are based on distance-based and prototype-based models, I will argue for a vital property of such models, namely components of their inherent interpretability, and I will explain exemplary applications from the domain of driver assistance and biomechanics.
Neural networks have revolutionised domains such as computer vision or language processing, and learning technology is included in everyday�s consumer products. Yet, practical problems often render learning surprisingly difficult, since some of the fundamental assumptions of the success of deep learning are violated. As an example, only few data might be available for tasks such as model personalization, hence few shot learning is required. Learning might take place in non-stationary environments such that models face the stability-plasticity dilemma. In such cases,
applicants might be tempted to use models for settings they are not intended for, such that invalid results are unavoidable.
Within the talk, I will address three challenges of machine learning �in the wild�:
(i) How to learn reliable given few examples only?
(ii) How to learn incrementally in non-stationary environments where drift might occur?
(ii) How to enhance machine learning models by an explicit reject option, such that they can abstain from classification if the decision is unclear
I will present robust approaches how to address these problems, which are based on distance-based and prototype-based models, I will argue for a vital property of such models, namely components of their inherent interpretability, and I will explain exemplary applications from the domain of driver assistance and biomechanics.
