Computer Vision News - August 2018

The case of pre-training 8 networks and using up to 4 of those networks for transfer-learning for each problem seems the most likely to come up in practice - - it is marked in red and zoomed on the right. An interesting result appears in the right column -- where all 26 networks were pre-trained -- there was nevertheless improvement in performance, indicating that transfer learning can be useful even when you have a network already trained on a massive labeled dataset which is relevant to your problem. Researchers also compared Gain and Quality : Gain measures how much was gained through transfer learning -- defined as the win rate of the pre-trained network using transfer-learning on the target task dataset vs. a network trained from scratch on the target task dataset. Quality measures performance vis-a-vis state of the art target task networks -- defined as the win rate against a network which was trained fully-supervised on a massive labeled dataset relevant to the target task. Research 9 Research Computer Vision News