Randomized dimensionality reduction of deep network features for image object recognition
- Publication Type:
- Conference Proceeding
- Citation:
- Proceedings - 2018 2nd International Conference on Recent Advances in Signal Processing, Telecommunications and Computing, SIGTELCOM 2018, 2018, 2018-January pp. 136 - 141
- Issue Date:
- 2018-03-26
Closed Access
| Filename | Description | Size | |||
|---|---|---|---|---|---|
| 08325778.pdf | Published version | 258.86 kB |
Copyright Clearance Process
- Recently Added
- In Progress
- Closed Access
This item is closed access and not available.
© 2018 IEEE. This study investigates data dimensionality reduction for image object recognition. The dimensionality reduction was applied to features extracted from an existing pre-trained Deep Neural Network (DNN) structure, the AlexNet. An analysis of the neurons in different layers of the AlexNet revealed an incremental increase in the pair-wise orthogonality between weight vectors of neurons in each layer, towards higher-level layers. This observation motivated the current study to evaluate the possibility of performing randomized dimensionality reduction by mimicking the observed orthogonality property of the high-level layers on activations of low-level layers of the AlexNet. Image object classification experiments have shown that the proposed random orthogonal projection method performed well in multiple tests, consistently outperforming the well-known statistics-based sparse random projection. Apart from being data independent, the proposed approach achieved performances comparable with the state-of-the-art techniques, but with lower computational requirements.
Please use this identifier to cite or link to this item: