An innovative design of neural network can be a solution to many challenges in AI

Dec 24, 2018News

David Duvenaud – an AI researcher in the University of Toronto and his collaborators at the university and the Vector Institute design a brand-new prototype neural network that can overcome its previous models.

At first, his idea was to create a deep-learning algorithm that could predict a person’s health over time. However, the data given by medical’s record is a bit complicated since each check-up gives different record with different reasons and measurements. Conventional machine-learning method finds it struggling to model continuous processes, especially those are not measured often. Because this method finds the patterns in data by stacking layers of simple computational nodes, the discrete layers keep it from providing the exact outcome. To be more specific, traditional machine-learning model follow its common process, known as supervised learning, which means collecting a lot of data’s layer to figure out a formula to solve other issues with similar traces. For instance, it mistakes a cat for a dog due to the fact that they both have floppy ears. However, there are various types of dog and cat with diverse features. Hence, it might produce inaccurate results.

In response to the difficulty, they allowed the network find formulas match the description of each stage of the process, each stage represents a layer of data. Taking the example of differentiating the two pets above, the first stage might take in all the pixels and use a formula to find out which ones are most similar for cats versus dogs. A second stage might use another to construct larger patterns from groups of data to tell whether the picture has whiskers or ears. Each subsequent stage would identify a feature of the animal, after collecting a sufficient data of layers, it will identify the animal’s picture. This step-by-step breakdown of the process allows a neural net to build more sophisticated models in order to produce a more accurate prediction.

Yet, in terms of the medical field, it will require us to classify health records over time into discrete steps for instance period of years or months. So, the only way to model these medical records is to specify it even more, it might encounter the same problems as the traditional model does. To make actual breakthroughs, they still need to dig deeper into the method with more experiments and research.

“The paper will likely spur a whole range of follow-up work, particularly in time-series models, which are foundational in AI applications such as health care,” said Richard Zemel, the research director at the Vector Institute.

No matter how far the algorithm has been advanced, there is still further risk that the rate at which AI advances will outpace the continuing development of ethical and regulatory frameworks. Layer 4 of the AIWS 7-Layer Model developed by MDI focuses on policies, laws and legislation, nationally and internationally, that govern the creation and use of AI and which are necessary to ensure that AI is never used for malicious purposes.