An unsupervised generative strategy for detection and characterization of rare behavioural events in mice in open fiel to assess effect of optogenetic activation of serotonergic neurons in the dorsal raphe nuclei

• Autor(a) principal: Soares, Lucas B. Nicolosi
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10451/40873
• topic: Redes neurais (Fisiologia); Aprendizagem automática; Controlo electrónico do comportamento; Serotonina; Inteligência artificial; Teses de mestrado - 2019; Domínio/Área Científica::Ciências Sociais::Psicologia; Domínio/Área Científica::Ciências Médicas

The purpose of our work is to provide an unsupervised deep learning tool that uses predictability of behavior as a meaningful metric to quantify the di erences between normal and abnormal behavior in the context of an experiment where mice receive optogenetic stimulation in their serotonergic neurons located in the dorsal raphe nuclei. We use generative adversarial networks to learn, on a training subset of the videos, a baseline behavioral repertoire by predicting future frames from subsequent frames in the past. By de ning a predictability index as dissimilarity between the quality of the generated prediction and the ground truth frame, we are able to determine in which frames a behavior not observed by the model during training is performed and therefore, we can detect the presence of stimulation by only analysing the uctuations of this index that indicate when the mouse is performing behaviors that are not present in the learnt baseline.