3D convolution neural networks (CNNs) have shown excellent predictive performance on tasks such as action recognition from videos, weather forecasting, detecting action similarity between two video clips, video captioning, labeling and surveillance. Also, they are used for performing object segmentation in 3D medical images. For example, in a 2D scan, vessels, bronchi and lung nodule all appear to be circular. However, a 3D scan can distinguish between a nodule, which is a spherical object, and a vessel, which is a cylindrical object.
Since 3D CNNs have huge computation and memory overheads, and have unique characteristics, custom accelerators are required for them. For example, training a 3D CNN on UCF101 dataset takes 3 days and training it on Sports-1M dataset takes nearly 60 days.
We review 30+ papers on hardware accelerators and hardware-aware algorithmic optimizations for 3D CNNs. These accelerators are for computing platforms viz., CPU, GPU, DSP, FPGA, etc. The algorithmic optimizations include pruning, use of FFT and Winograd based convolution, etc.
Paper is here, accepted in Journal of Systems Architecture, 2021.
Since 3D CNNs have huge computation and memory overheads, and have unique characteristics, custom accelerators are required for them. For example, training a 3D CNN on UCF101 dataset takes 3 days and training it on Sports-1M dataset takes nearly 60 days.
We review 30+ papers on hardware accelerators and hardware-aware algorithmic optimizations for 3D CNNs. These accelerators are for computing platforms viz., CPU, GPU, DSP, FPGA, etc. The algorithmic optimizations include pruning, use of FFT and Winograd based convolution, etc.
Paper is here, accepted in Journal of Systems Architecture, 2021.