L2S Talk by Leonardo Rossi Luiz (University of São Paulo): Adding Selective State Space Models to Deep Learning Architectures for Human Activity Recognition
Thu 18 November 2026, 16.00 h at H-A 8107
Continuous wearable monitoring has made Human Activity Recognition (HAR) essential for intelligent systems. While current deep learning models based on sliding windows excel at classifying short and isolated data segments, they often fail to capture long-range temporal dependencies and inter-window contexts. In that regard, this research proposes the adaptation of the DeepConvContext architecture by integrating Selective State Space Models (SSMs) to replace traditional Recurrent Neural Network (RNN) layers. By leveraging the ability of SSMs to dynamically select and process input sequences, this project aims to overcome the fixed-gate limitations of LSTMs while avoiding the high computational complexity of Transformers.

