Modern neural networks are powerful predictive models. However, when it comes
to recognizing that they may be wrong about their predictions and measuring the
certainty of beliefs, they perform poorly. For one of the most common activation
functions, the ReLU and its variants, even a well-calibrated model can produce incorrect
but high confidence predictions. In the related task of action recognition, most
current classification methods are based on clip-level classifiers that densely sample a
given video for non-overlapping, same sized clips and aggregate the results using an
aggregation function - typically averaging - to achieve video level predictions. While
this approach has shown to be effective, it is sub-optimal in recognition accuracy
and has a high computational overhead. To mitigate both these issues, we propose
the confidence distillation framework to firstly teach a representation of uncertainty
of the teacher to the student and secondly divide the task of full video prediction
between the student and the teacher models. We conduct extensive experiments
on three action recognition datasets and demonstrate that our framework achieves
state-of-the-art results in action recognition accuracy and computational efficiency. / Thesis / Master of Science (MSc) / We devise a distillation loss function to train an efficient sampler/classifier for video-based action recognition tasks.
Identifer | oai:union.ndltd.org:mcmaster.ca/oai:macsphere.mcmaster.ca:11375/26127 |
Date | January 2020 |
Creators | Manzuri Shalmani, Shervin |
Contributors | Chiang, Fei, Zheng, Rong, Computing and Software |
Source Sets | McMaster University |
Language | English |
Detected Language | English |
Type | Thesis |
Page generated in 0.0018 seconds