Learning from Data (Fall 2022)

[kundtx]

http://8.129.175.102/lfd2022fall-poster-session/3.html

[yang2011]

(Instructor) Great visual representation! What does the x-axis label represent in the bar plot and the bottom line plot in the right column?

[NineAbyss]

G19 Peisong Wang: Impressive work! What are the novel classes and how much data do you have for the novel classes?

[uprightman47]

@yang2011 (Instructor) Great visual representation! What does the x-axis label represent in the bar plot and the bottom line plot in the right column?

G3 Siqi Chen: Thank you, professor! In the bar plot, the x-axis label represents the number of different settings of target classes. In our experiments on Xsens DOT dataset, a portion of the classes is selected as the source classes and the remaining as the target classes. We have tried four kinds of target classes settings. And the figure shows that the best performances are almost always achieved by FSHAR methods. The cross-domain class-wise relevance measure embedded in our framework is helpful to improve the few-shot learning performance.

The x-axis label of bottom line plot in the right column is the size of the hidden layers in LSTM. We find that too small or too large size can result in poor performance. So the experiments are designed to analyze the impacts of the size of the hidden layers.

[ErlindaQiao]

@NineAbyss G19 Peisong Wang: Impressive work! What are the novel classes and how much data do you have for the novel classes?

G3 Xizi Qiao: Thanks for your question! 9 kinds of activities in total are divided into two parts: source domain ( includes 6 activities) and target domain ( includes 3 activities). The novel classes refer to the activities in the target domain, the ones with few available data and also the ones we want to classify. The target domain contains 2922 pieces of original data recorded by sensors.

[ILoveCodeP]

G16 Li Jipeng :I would like to ask what are the main applications of this behavioural classification in our lives? As the data appears to be collected by a wearable device, I am wondering if this can only be collected in an experimental setting and not applied to everyday life?

[NineAbyss]

@ErlindaQiao

@NineAbyss G19 Peisong Wang: Impressive work! What are the novel classes and how much data do you have for the novel classes?

G3 Xizi Qiao: Thanks for your question! 9 kinds of activities in total are divided into two parts: source domain ( includes 6 activities) and target domain ( includes 3 activities). The novel classes refer to the activities in the target domain, the ones with few available data and also the ones we want to classify. The target domain contains 2922 pieces of original data recorded by sensors.

Thank you! Your model seems effective:)

[ErlindaQiao]

@ILoveCodeP G16 Li Jipeng :I would like to ask what are the main applications of this behavioural classification in our lives? As the data appears to be collected by a wearable device, I am wondering if this can only be collected in an experimental setting and not applied to everyday life?

G3 Xizi Qiao: Well, actually it has a wide range of application in reality. For example, in the field of health care, sensors are used to track the motion of patients during the surveillance or recovery phase. Common motions, such as walking and sitting, already have abundant training data, thus can be recognized easily. But uncommon or personalized motions, like starting to walk with aid, are different from normal motions and are critical to be recognized. Consequently, human activity recognition models which can detect activities with very few training samples are needed.

[ErlindaQiao]

@NineAbyss

@ErlindaQiao

@NineAbyss G19 Peisong Wang: Impressive work! What are the novel classes and how much data do you have for the novel classes?

G3 Xizi Qiao: Thanks for your question! 9 kinds of activities in total are divided into two parts: source domain ( includes 6 activities) and target domain ( includes 3 activities). The novel classes refer to the activities in the target domain, the ones with few available data and also the ones we want to classify. The target domain contains 2922 pieces of original data recorded by sensors.

Thank you! Your model seems effective:)

Thanks for comments! :D