Abstract
Deep learning approaches are now a popular choice in the field of automatic emotion recognition (AER) across various modalities. Due to the high costs of manually labeling human emotions however, the amount of available training data is relatively scarce in comparison to other tasks. To facilitate the learning process and reduce the necessary amount of training-data, modern approaches therefore often rely on leveraging knowledge from models that have already been trained on related tasks where data is available abundantly. In this work we introduce a novel approach to transfer learning, which addresses two shortcomings of traditional methods: The (partial) inheritance of the original models structure and the restriction to other neural network models as an input source. To this end we identify the parts in the input that have been relevant for the decision of the model we want to transfer knowledge from, and directly encode those relevant regions in the data on which we train our new model. To validate our approach we performed experiments on well-established datasets for the task of automatic facial expression recognition. The results of those experiments are suggesting that our approach helps to accelerate the learning process.
Highlights
In recent years deep learning approaches, have become a popular choice in the field of automatic emotion recognition
Inspired by findings that the human learning process can be accelerated by providing visual cues (Mac Aodha et al, 2018) we aim to use those saliency maps to guide the attention of a Trainee network, which we want to focus on areas that were relevant for another model that has been pretrained on a similar task
Since we have shown that our Relevance-based Data Masking (RBDM) approach, in its current form, can achieve a positive impact on the training speed but does not contribute much to increase the overall performance of the Trainee model we think that the choice of transfer learning method depends on the intended application
Summary
In recent years deep learning approaches, have become a popular choice in the field of automatic emotion recognition. Convolutional Neural Network (CNN) architectures are promising to overcome the limitations of handcrafted features by directly learning suitable representations from raw data. To train a CNN that handles such raw data input with high accuracy from scratch, vast amounts of annotated data are necessary as the absence of handcrafted features requires additional abstraction layers to be automatically learned by the network. A common solution to this problem comes in the form of transfer learning where the previously gained knowledge of a model about related tasks is used to facilitate the learning process. There are three potential benefits that can be gained from transfer learning (Torrey and Shavlik, 2010): Increased initial performance of the new model, a steeper learning slope that leads to faster learning and an increase in the final performance level that the model can achieve
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
