MCL is joining the research of Convolutional Neural Network (CNN). With the donation of one K40 GPU from NVIDIA, MCL will create its first GPU server to train and test various CNN architectures. K40 is the prestigious GPU that is widely used in parallel computing environment. It features 12 GB memory and 2880 CUDA cores that can significantly accelerate the speed of modern computer vision algorithms. When ECC is turned off and clock frequency is slightly boosted, it is capable of processing 20 iterations of Caffe training in 19.2 seconds.
MCL director, Prof. C.-C. Jay Kuo, appreciates NVIDIA’s generous donation and has assigned Phd students, Hao Xu and Qin Huang to build a powerful server to house the K40. Prof. C.-C. Jay Kuo is interested in solving various computer vision problems. He has a vision that there should be a systematic way to provide unified solution to all types of computer vision problems. To achieve that goal, Prof. C.-C. Jay Kuo believes that we need to better understand the strength of the automatically trained feature sets.
MCL has a detailed road map for its future development of Convolutional Neural Network, and one crucial step is to study the features trained through it. MCL will use Caffe and Theano libraries to test various CNN architectures, from the classic AlexNet to the deeper and more accurate VGG net. The visualization of features trained from these architectures will be carefully examined using deConv network. Finally, MCL wants to build a clear understanding towards both the automatically trained features themselves and the reason for why are they selected by the CNN.