Brief Description
This video was created as a part of the Topics in Machine Learning course at UCF under Dr. Ulas Bagci. This video covers the concept of why deep, wide networks tend to achieve high performance, yet 90% of the parameters can be removed after training and still achieve the same performance level. However, if we re-train that pruned network from scratch, the performance will significantly decrease. So what's going on here!? Well it might be something called "The Lottery Ticket Hypothesis", checkout the video! The original work was published in ICLR 2019 winning the best paper award and was written by Jonathan Frankle and Michael Carbin from MIT CSAIL. If you enjoyed this video, you can view more videos from the course webpage.
Follow up
If you enjoyed this video please let me know. I am considering starting a dedicated channel where I explain papers in this level of detail on a regular basis.