Keynote | Technical
Friday 17th | 14:00 - 14:30 | Theatre 20
Keywords defining the session:
- Machine learning
- Deep nets
Time series related problems have traditionally been solved using engineered features obtained by heuristic processes. Based on the recent success of recurrent neural networks for time series domains, we present a generic deep learning framework for time series based on convolutional and LSTM recurrent units, which: (i) works on raw temporal data; (ii) does not require expert knowledge in designing features; and (iii) explicitly models the temporal dynamics of features. We detail the architecture and present an actual implementation of such framework. We illustrate the performance of the framework on a use case: the problem of human activity recognition.