Conclusion

In this course, we have focused on learning from time series and we have detailed two families of methods.

The first one relies on matching-based metrics that aim at assessing similarity between time series through temporal alignment. In this context, we have focused on the Dynamic Time Warping (DTW) method and its differentiable counterpart softDTW.

In a second part, we have focused on neural network architectures tailored to deal with time series, and more specifically convolutional models as well as recurrent ones.

Finally, we have seen that methods from both parts of the course could be combined, through the use of differentiable alignment-based similarity measures as loss functions when time series are provided as model outputs.

Lab sessions

The lab sessions for this course are available on GitHub at rtavenar/notebooks-ml4ts.