Depends on the specific appliaction you plan to use. There are three advantages to Threadripper (HEDT):
More cores, current TR goes up to 64 cores
More memory channels, allows for more total memory and more bandwidth
More PCIe lanes, mainstream has 24 on the CPU TR has 64
For Deep Learning TR would probably be the way to go honestly. DL tends to take advantage of GPGPU so more high performance GPUs would be beneficial and more PCIe lanes would help that.
However Intel is showing off a DL Boost which they say is supported in their 10th gene CPUs. I am sure we will see more in the next few years from all sides though.