Machine Teachers for Active Learning Deep Learning
- Delivery
- Available on this site
- Format
- Price
- Non-members (tax incl.):¥1,100 Members (tax incl.):¥880
- Publication code
- 20219026
- Paper/Info type
- Other International Conferences
- Pages
- 1-4(Total 4 p)
- Date of publication
- Sep 2021
- Publisher
- JSAE
- Language
- English
Detailed Information
Author(E) | 1) Mohd Hafiz Hilman Mohammad Sofian, 2) Toshio Ito, 3) Yasutaka Okada, 4) Hiromi Rei, 5) Ogishima Aoi |
---|---|
Affiliation(E) | 1) Shibaura Institute of Technology, 2) Shibaura Institute of Technology, 3) Denso Ten Limited, 4) Denso Ten Limited, 5) Denso Ten Limited |
Abstract(E) | Deep Learning (DL) has enabled people around the world to create a reliable object detection system. It is now being used in developing an autonomous vehicle that can detect and understand its surroundings through the camera. That being said, it is hard to train a reliable deep learning based object detection system since a lot of data is needed to be annotated. This introduces high cost of time and human power. This problem can be solved through Active Learning (AL) framework, where the model itself chooses a subset of data that it wants to be trained on and human just need to annotate that particular subset of data. This method cannot just reduce the cost of time and human power, but also the total time of learning. However, human interference is still needed in the data annotation process. This paper proposed Machine Teachers (MTs), an algorithm that is designed to be used by AL framework to automate the process of data annotation. We named the combination MTs and AL framework as Automated Active Learning (AAL) framework. MTs is essentially an algorithm that incorporated an already trained deep learning model with conventional object detection methods in order to automatically annotate the data based on feedback from the AL framework. MTs can be improved and adjusted based on the type of data and object class to be annotated, making it flexible for a lot of usage. Our early experiment on MTs performance while comparing it to the YoloV3 model on the precision matrice shows that there are potential for MTs to really be a valuable method in making AL a more effective training method. |