naturewn.com

Trending Topics NASA cancer dinosaur ISS International Space Station

Researchers Develop New AI Algorithm That Can Learn Simply by Human Instruction

Nov 18, 2016 06:33 AM EST
WATCH RELATED VIDEO
Researchers Develop New AI Algorithm That Can Learn Simply By Human Instruction
A new machine learning algorithm from engineers at U of T Engineering may soon be able to allow users to directly instruct AI without programming.
(Photo : Photo by Stephen Lam/Getty Images)

A new machine learning algorithm from engineers at U of T Engineering may soon be able to allow users to directly instruct AI than an existing set of examples. 

U of T Engineering researchers Parham Aarabi and Wenzhi Guo designed a special algorithm that can learn directly from human instructions. This has outperformed conventional methods of training neural networks by 160-percent. Interestingly, the new algorithm also outperformed its own training by nine percent. 

According to Phys.Org, the new algorithm has learned to recognize hair in pictures with greater precision than that enabled by the training. This marks a rather significant leap in the field of artificial intelligence. 

Aarabi and Guo have trained their algorithm to be able to identify people's hair. This is a task proven difficult for computers, given that the texture of the background can be misconstrued as hair. Aarabi explained the algorithm's progress is akin to like a teacher instructing a child, and a child eventually learning beyond what was taught to him or her.

Humans can "teach" networks by providing a set of data and asking the network to make decisions based on the samples it has seen. However, the new method learns directly from human trainers. Called "heuristic training," humans can provide direct instructions used to pre-classify training samples.

This includes instructions such as, "Sky is likely to be varying shades of blue" or "Pixels of a sky are likely to be found at the top of the image instead of below."

The new method holds huge potential in advancing the field of AI. This is also a potential solution to one of the biggest challenges for neural networks, which is to make correct classifications of previously-unknown or unlabeled data. This is important to be able to help machines learn from new situations. Applicatons include correctly identifying cancerous tissues or classifying objects. 

© 2017 NatureWorldNews.com All rights reserved. Do not reproduce without permission.

Join the Conversation

arrow
Email Newsletter
About Us Contact Us Privacy Policy Terms&Conditions
Real Time Analytics