Tuesday, November 13, 2018
Sunday, November 11, 2018
Tuesday, October 30, 2018
Tuesday, October 23, 2018
We are all lying to ourselves
"On any given day, 80% of us are lying to ourselves about lying to ourselves", Tasha Eurich
Sunday, October 21, 2018
Machine Learning is artificial, but it's not necessarily intelligent
Machine Learning is "artificial", but it's not really intelligent from a certain point of view. There appears to be a trend going on in the industry right now that uses the terms "Machine Learning" and "Artificial Intelligence" interchangeably (or at least sees Machine Learning is a type of AI).
Intelligence is the often measured as the ability to see connections between different things. Machine Learning doesn't see connections. It just uses CPU time for brute-force analysis to find connections based on many iterative cycles of failures and successes in stages. Early failures are often permanently disregarded, and early successes given too much value, even if they follow deadend paths . Machine Learning typically is not able to see that future success could actually come from pathways where it experienced early failure, unless there is human intervention of some sort.
This video should scare all of us, and not because of the wolves:
Amazon learned a similar lesson on its own regarding Machine Learning, as detailed in the following venturebeat.com article. The article describes how Amazon had to scrap their Machine Learning program for hiring people. The program's purpose was to remove gender bias in hiring of new employees. However, their program developed gender bias on its own, despite the development team's best efforts to remove bias. Article: Amazon scrapped a secret AI recruitment tool that showed bias against women [archive.org].
This article on Volt DB (6 Reasons Why Your Machine Learning Project Will Fail to Get Into Production [archive.org]) goes into common problems with Machine Learning projects. It boils down to data quality. The problem is, data from the real world will always be of poor quality.
Intelligence is the often measured as the ability to see connections between different things. Machine Learning doesn't see connections. It just uses CPU time for brute-force analysis to find connections based on many iterative cycles of failures and successes in stages. Early failures are often permanently disregarded, and early successes given too much value, even if they follow deadend paths . Machine Learning typically is not able to see that future success could actually come from pathways where it experienced early failure, unless there is human intervention of some sort.
This video should scare all of us, and not because of the wolves:
This article on Volt DB (6 Reasons Why Your Machine Learning Project Will Fail to Get Into Production [archive.org]) goes into common problems with Machine Learning projects. It boils down to data quality. The problem is, data from the real world will always be of poor quality.
Sunday, October 14, 2018
Subscribe to:
Posts (Atom)