Technology

AI Algorithm Surpasses What it Was Taught by Humans

artificial intelligence

U of T Engineering researchers  Wenzhi Guo (ECE MASc 1T5) and Parham Aarabi (ECE) have designed a new machine learning algorithm that may soon enable your smartphone to give you an honest answer based on logic.

The algorithm does not learn from an existing set of examples, but rather takes its data directly from human instructions. This methodology resulted in it outperforming conventional methods of training neural networks by a whopping 160 per cent.

What is even more surprising is that the algorithm also outperformed its own training by nine per cent.  It learned, for example, to recognize hair in pictures with more reliability than that enabled by the training. This result is heralded as a significant leap forward for artificial intelligence.

Guo and Aarabi taught their algorithm to use photographs to identify people’s hair. This task is much more challenging for computers than it is for humans.

Aarabi explains that their algorithm managed to classify difficult, borderline cases correctly by differentiating between the textures of the background versus the texture of hair. He compares this to a teacher instructing a student, and the student learning beyond what the teacher taught her initially.

This figure compares a traditionally trained algorithm to Aarabi and Guo’s heuristically trained neural net. The left and centre columns show an aggressive and conservative image-recognition algorithm trained to recognized human hair, compared to the more precise heuristically trained algorithm at right. (Image credits: IEEE Trans NN & LS)
This figure compares a traditionally trained algorithm to Aarabi and Guo’s heuristically trained neural net. The left and centre columns show an aggressive and conservative image-recognition algorithm trained to recognized human hair, compared to the more precise heuristically trained algorithm at right. (Image credits: IEEE Trans NN & LS)

Humans normally teach computer networks that learn dynamically (neural networks) by providing a set of labeled data. The neural network is then asked to make decisions based on the samples it’s been shown. You could for example show a neural network hundreds of pictures with the sky labeled, to train it to identify sky in a photograph.

Guo and Aarabi’s algorithm is different in that it learns directly from human trainers. This model is called heuristic training and rather than showing a set of fixed examples, humans make direct instructions that are used to pre-classify training available. The algorithm is also provided with guidelines such as telling it that pixels near the top of the image are more likely to be sky than pixels at the bottom, and the sky is likely to be various shades of blue.

Making correct classifications of previously unlabeled or unknown data is one of the biggest challenges facing neural networks. This heuristic training approach holds significant promise for removing this obstacle.

This is critical for applying machine learning to unknown situations, such as classifying all the objects approaching and surrounding a self-driving car, or correctly identifying cancerous tissues for medical diagnostics.

Guo is keen to apply their method to a range of applications and other fields, from transportation to medicine. He also notes that supplying heuristic training to hair segmentation is just the beginning.

The study has been published in the journal IEEE Transactions on Neural Networks and Learning Systems.

Save