Conflict of interest

domain, the small rate difference (2.21%) between "TL SVM across domains" and "TL SVM similar domains" suggests that the low- and mid-level feature space generated by human activities shares common details with the one generated by movements of specific atypical subjects. This is not the case for frequency-domain series, which can be explained by the difference in the frequency range between human activities and SMMs. Indeed, the FFT

amplitude of human activities is contained below 10 Hz, pre-training the CNN on human activity frequency-signals from 0 to 3 Hz and not from 0 to 10 Hz results in imperfect human activity features which, combined with the SVM, do not seem to yield good classification results on the recognition of SMMs. If we were to have a new target learning task whose data signals are within the same frequency range as data signals of the source learning task, then "TL SVM across domains" would have achieved the same performance as "TL SVM

• One advantage of "TL SVM similar domains" and "TL SVM across domains" is that they can be implemented in Android portable devices, as shown in Table 4. Indeed, an expert could receive continuous acceleration signals from the torso accelerometer of a subject, and label them on the fly (as SMM/non-SMM) as the subject performs his activities/movements. This results in annotated time series which are then preprocessed and fed into either "TL SVM similar domains" or "TL SVM across domains" for training. A one-minute recording of these signals is sufficient to train one of the two frameworks. Afterwards, this framework is ready to use for recognizing further SMMs on

Time series pose important challenges to existing approaches which perform predictive modeling for classification tasks. In this paper we present a review on our previous works. Our contributions are aggregated into two categories: data-level and algorithm-level approaches. Our data-level approach consists of encoding time series using STin order to produce noise-free input signals which offer a more efficient CNN training. At the algorithm level, one approach is the adaptive convolutional layer filter approach which consists of determining the size of the filter based on an analysis of the input time series signals and fluctuations present within them. Indeed, choosing the proper 1st layer filter generates features maps which are more informative about the input signals and which capture the whole peaks within input signals. Furthermore, "TL SVM similar domains" and "TL SVM across domains" are algorithm-level approaches dealing with tasks with limited annotated data, which are regarded as two global, fast and light-weight techniques for these kinds of tasks. These two CNN approaches generate features general and global enough to recognize time series of the target learning task, given time series of a source learning task that is similar or different but related to the target learning task. All these approaches were implemented on the recognition of human activities, including normal activities performed by typical subjects and disorder-based activities performed by atypical subjects (such as SMMs of autistic subjects). Experimental results have showed the superiority of our techniques and their ability to extract relevant features from time series inputs. As a perspective, knowing that

time series datasets often contain outliers either due to noisy time series or

mislabeled time series (e.g. incorrect labels), we aim at studying a robust CNN that is insensitive to outliers. As opposed to our data-level CNN technique (mentioned in

similar domains".

Time Series Analysis - Data, Methods, and Applications

that same subject.

5. Conclusion

76

The authors declare that they have no conflicts of interest.
