How can one expect a biometric system that relies on a single enrolment sample to cope with all the variability possibly encountered during the operational phase? The key idea is to make the system adaptive in the context of multimodal biometrics. Multimodal biometrics is about combining multiple biometric traits and it can be shown theoretically that this approach leads to improved recognition accuracy. This lecture will explore some aspects of multmodal biometric adaptation, ranging from the use of quality measures, user-specific statistics and cohort information to the new exciting development of template-update and adaptive threshold/score normalization techniques. 

The goal of the lecture is to show how the above problems can be solved using machine learning techniques as fundamental building blocks. For instance, it will be shown how a clustering algorithm can be combined with a generative/discriminative classifier to form a mixture of linear classifiers that results in the state-of-the-art classifier for quality-based fusion. Another example is how the cohort information can be modelled first by regression and then solved as a classification problem. 

The main topics to be covered are:

• quality-based fusion
  
• client-specific fusion
• cohort-based score normalisation