At AIxIA, Cameron Schuler from the Vector Institute will give a keynote on Artificial Intelligence and IoT. Here’s a little sneak peek of what to expect.
Hi Cameron! At AIxIA you’re going to give a keynote on „How Reinforcement Learning can be applied to IOT“. Before we dive into the subject, can you tell us a bit more about yourself?
My career has covered business development, consumer products, finance, IT, and general management from start-ups to mature companies. I’ve even spent eight years in investment sales and trading.
I’ve been involved in the domain of AI and machine learning since 2008, when the field was basically considered to have no economic value and was pretty small. I joke that it sounded like a great place to accelerate my career – but as it turns out there is value.
I was one of the authors of the Pan-Canadian Artificial Intelligence Strategy and Canada was the first country to have a national AI strategy. For eight years I led Amii (amii.ca), one of the top-ranked Machine Learning and AI groups in the world. And now I’m with the Vector Institute, which was launched in 2017 as an independent, not-for-profit corporation dedicated to research in the field of artificial intelligence (AI).
On the website of Vector Institute one of the first sentences is about AI fostering economic growth and improving the lives of people. Do you have some examples for us, how this can be achieved?
Of course! We have 28 large companies that we call „enterprise sponsors“ and we have another 19 startups and scaleups. There are two reasons why we have sponsors: One is because it’s important to help grow the ecosystem and support it. And the second is to provide benefits to our sponsors.
Let me give you some examples. When you think about how we grow talent, experiential learning is a big piece of that. So, we worked with our sponsors on transformer-based NLP models to create the base level of knowledge and proficiency, and then the companies go back and apply it to their own domain.
An example of this is the Bank of Montreal (BMO). Together with our institute it used several online financial news sources to add over 182 million finance and market-related terms and their contexts to a data set. After that they pre-trained a model with this enriched dataset to achieve specific tasks for BMO on analyzing market sentiment.
Another example is Manulife Insurance. Its data scientists focused on topic modelling, a technique that automates document classification according to topic. As an insurance company, Manulife has copious amounts of unstructured text (customer call transcripts, benefit submissions, etc.) and automating the accurate reading, analysis, and sorting of these documents is incredibly valuable for them. The goal here is to enable the company to generate value by better servicing their customers through AI.
One last example that’s probably quite interesting: AI models were and are trained using historical data. The challenge becomes what happens when the world changes either slowly over time or a sudden event like a financial crisis or pandemic. The historical data doesn’t lose its value, but it is no longer useful for forecasting. In other words: Past performance doesn’t guarantee future results.
When the COVID pandemic hit, we started a project called Dataset Shift. It was all about better understanding of dataset shift principles, strategies for detecting shift, and techniques for adapting it. When put into production, this expertise and new advanced toolbox can enable us to maximize the effectiveness and resilience of AI systems, even during times of sudden and severe external change.
As already mentioned, your keynote will be about reinforcement learning and how it can be applied to IOT. For those who are not familiar with the matter, can you explain how Reinforcement Learning works?
Rich Sutton, whom I worked with for 10 years, wrote the textbook on Reinforcement Learning – and the basic methodology he uses when speaking about this technique is „How do we learn as humans?“
If we have bad behavior, we get penalized. If we have good behavior, we get rewarded. In other words, Reinforcement Learning is a type of machine learning technique that enables an agent to learn by trial and error using feedback from its own actions and experiences. It discovers its world based on certain predefined goals and will figure out what is needed to reach these goals.
Reinforcement Learning is quite hard in the real world, but ideally suited for controlled systems, because it’s about decision making.
How do AI and IoT fit together?
Historically systems are highly engineered, which requires a lot of people and they don’t have the ability to change over time. It’s all hard coded. What AI allows us to do is to learn from the systems and help them adapt. Of course, it’s going to be more complex than that, because ultimately there are few if any models in the real world that learn and are deployed in real time. Usually, you effectively train them and once you understand how the model works, you deploy them.
Our world is increasingly complex, especially computing systems. This makes learning systems important. First, it can assist when things are too complex, and you don’t know how to get from one end to the other. Second, things change over time and the question is how to adapt to that change. Third, you know where you are and you know where you want to go, but you don’t know how to get there.
In your opinion, where are the limits of artificial intelligence, if any?
Well, the larger the problem gets, the larger the model gets and more importantly the greater the computing requirements – and there’s only limited computing capacity in the world. Let’s look at games. Checkers is the largest game that has been solved to date, with total possible moves of 5×1020. If you look at large games like poker, you can’t solve them. And then of course there’s a difference between a game like poker which is an incomplete information game (at best you can get 15 percent of the information) and a game like chess where you can see the whole board. All possible move variations in the game of chess are estimated between 10111 and 10123 – and that is not only unsolvable but there is no clear pathway for the required computing to solve that.
If you think about even bigger problems, the models and computing we need grow exponentially larger. The real world is hard to work with and there are still many limitations.
There is no part of our lives and future generations that won’t be impacted by AI. In the field of machine learning we are trying to make the best decisions in ambiguous environments. We as humans think we make rational decisions. We try to make the best decision given the information provided. And machine learning works the same way.
About the project:
The project „French-German Networking on Industry 4.0 and Applied Artificial Intelligence“ is supported by the Ministry of Economic Affairs, Labour and Tourism Baden-Württemberg and coordinated by an experienced consortium, consisting of DIZ | Digital Innovation Center, Allianz Industry 4.0, bw-International, CyberForum e.V, FZI | Research Center for Information Technology and Steinbeis Europe Center . The main goal of the project is to lay down a foundation for a sustainable French-German ecosystem on AI and Industry 4.0. More informations about the French-German Project in french or german.