OSHEAN eCurrent Fall 2019 | Page 26

algorithms and how our own visual system still would be hard to collect in practice). This is vastly outperforms state-of-the-art algorithms. part of the reason why companies currently I hope attendees walk away from my have self-driving cars covering hundreds or presentation understanding that AI is not yet thousands — if not millions of miles — with a solved and there is still a lot to learn from human behind the wheel. They are able to catch neuroscience. mistakes and continuously increase the image datasets used to train algorithms to include all What do you expect to see in the computer vision space in five years? edge cases. In five years, we are going to see more of what I believe we will also soon be facing a moral we are seeing today. debate in which human mistakes at the wheel will decrease with AI-assisted driving, but I think we will witness medical diagnosis semi-random mistakes made by computer scaling up to, and possibly beyond, human vision will also occur that are hard to explain. performance. Doctors will be able to leverage There is definitely going to be a learning curve; and collaborate with computer algorithms one that I’m excited to see play out. Major Internet Social Media Networks and When They Were Started Dialup in 1999 was 56KB/sec — August 1, 2003 — February 14, 2005 for diagnoses. — January 15, 2001 We’ll certainly start seeing the advent of more AI-based systems that are able to assist — March 21, 2006 humans. However, complete autonomy for self-driving vehicles, for example, is still a long way away because the systems are presently unable to generalize properly to #TECHTHROWBACK Spotlight Year 1999 — February 2004 novel situations in the same way humans would. Even though autonomous vehicles are pretty accurate — they can drive thousands of miles without problems — they will make mistakes every once in a while. For instance, an autonomous vehicle may miss a pedestrian crossing the road. The primary issue is the limited ability of computer vision algorithms to generalize beyond the data used to train them. For example, these algorithms Internet access speeds have grown from 56 KB/sec dial-up in 1999 to Gigabit-range broadband, with speeds up to 100 GB/sec for high-bandwidth applications. — August 29, 1997 streaming content 2007 — 2002 — October 6, 2010 cannot detect a baby crossing the road unless they have been explicitly trained to (which 26 | 2019 eCURRENT Broadband in 2019 is 100GB/sec Stronger Together | 27