The following kernels are useful for machine learning. They will benefit from being implemented at a lower level so that we can take advantage of backend specific kernels to accelerate them. cuDNN, MIOpen, and mkl-dnn, should help, for instance.

• StridedConvolve
• Pooling
• Batched Matmul (Related: 3D and 4D support for Linear Algebra functions #483)
• 8bit Matmul
• Vectorized dot product