Recap: Workshop at the Ottawa Artificial Intelligence Alliance (OAIA)
Canada: the land of maple syrup, moose, hockey, but also, AI! Throughout the country there are many notable research communities, and the National Capital Region is no exception. To better exploit all the interesting work in government organizations, IT companies and universities in the area, the Ottawa Artificial Intelligence Alliance formed a scientific society, with founding members including Carleton University, uOttawa, and the National Research Council Canada. Their main goals consist of increasing visibility, networking opportunities, community development and involvement.
This month, in line with these goals, the alliance organized a workshop at the National Research Council featuring speakers from both academia and industry. Topics of talks included AI’s Successes and Challenges, Detecting Signs of Mental Illness from Social Media, Probabilistic Generative Deep Learning, as well as Computational Intelligence Techniques for 3D Object Modelling. In addition to these talks, the workshop provided the opportunity for local researchers to submit posters for presentation.
We were invited to host a talk regarding our unsupervised method of text embedding. The talk was presented by Hichem Mezaoui, PhD., one of our talented Machine Learning Researchers. This method is based on a generative model, referred to as the random walk model.
A common challenge in many industries is trying to cut costs. Companies can only afford to implement solutions if they help rather than hurt the bottom line. As such, cheap and effective is the way to go. The random walk model holds true to that need for a cost effective solution, outperforming more sophisticated deep neural net models such as RNN and LSTMs on different semantic tasks, such as textual similarity, entailment (which is a directional relation between elements of text) and sentiment analysis. With deep neural net models you’re looking at a model that is not only more complex to implement but requires more processing power and time to run. This results in higher costs that you don’t see with generative models.
The new proposed framework could be further extended to a wide variety of applications. Particularly, it has the capacity to address the main problems within the domain of natural language processing. These problems include text features embedding, word sense disambiguation and summarization.
At IMRSV Data Labs we are actively collaborating with partners in academia and industry alike to solve problems like these. If this is something that interests you or you would like to contribute to the efforts, feel free to reach out to us!