|
Classification of Ecological Data Using Deep Learning Methods
Presentation Number:0047 Time:15:52 - 16:04
|
Frank Y. Shih, Shaobo Liu, Gareth Russell, Kimberly Russell and Hai Phan
Deep learning methods have become increasingly important as the advancement of computing speed and capacity during recent years. When applied to computer vision and image processing, deep learning has great performance on object classification, object detection, image segmentation, etc. In this paper, we apply convolutional neural network to classify ecological data such as 19 types of bee wings. One difficulty in the ecology classification is the limited amounts of the dataset. Thus, image augmentation method is used to increase the dataset size. The original dataset of 750 images is enlarged to 19,000 images, among them 15,200 images used for training (i.e., 800 images for each class) and 3,800 images for testing. With the augmented images, we train our model to classify the types of bees according to their wings. Furthermore, we apply transfer learning technique to improve model’s classification accuracy. A pre-trained neural network model is applied on the ecological dataset. The test accuracy of 98.34% is achieved after data augmentation. And 93.79% test accuracy for trans-fer learning, which is used to build one-class classification and recognize “unknown” class.
|
|
|
|
|
Current Dipoles Analysis for Alpha Activity of Eeg Neurofeedback Training
Presentation Number:0164 Time:16:40 - 16:52
|
KEN-HSIEN SU, Jen-Jui Hsueh, Tainsong Chen and Fu-Zen Shaw
Neurofeedback training (NFT) of electroencephalogram (EEG) is a psychophysiological procedure that allows users to learn self-regulation of their cortical oscillations. Trained alpha activity after NFT is associated with cognitive performance, its neural origin, however, remains unknown. The present study aimed to explore possible equivalent current dipoles (ECDs) of the trained alpha activity. Thirty-six healthy participants were recruited and randomly assigned to either an Alpha group with feedback of 8-12 Hz (n=20) or a Sham group with feedback of 4-Hz bandwidth amplitude selected randomly from 7-20 Hz (n=16). The Alpha group exhibited progressively significant increase of 8-12-Hz amplitude exclusively. Furthermore, the Alpha group had reliable controllability of the trained alpha occurrence in a block deign. ECDs of the trained alpha activity existed in the posterior cingulate cortex, precuneus, middle temporal cortex, and hippocampal formation, which are associated with several cognitive functions. There was no ECD within the occipital cortex. These results suggest that the trained alpha activity differs from classic alpha rhythm and may play an active role in cognitive performance.
|
|
|