Uncertainty and Artificial Intelligence

artificial intelligence.png
6 MINUTE READ

I love a gal who calls it like she sees it. Especially one of the foremost scientists in the high hype, high promise world of artificial intelligence.

What we don’t have right now is systems that are able to solve many different tasks at once and can reason about how to solve them. This is what is referred to as artificial general intelligence, and not only do we not have it, we don’t really have a clear idea of how to get there.
— Inmar Givoni, Autonomy Engineering Manager at Uber Advanced Technology Group

Artificial general intelligence refers to computers being able to perform a broad range of unfamiliar intellectual tasks in a manner similar to humans. It is a precursor to the Singularity, when computer intelligence will vastly outstrip human intelligence. I heard Inmar speak, along with the founders of Mindbridge.ai and Integrate.ai, at an event organized by the students at Branksome Hall, an independent girls' school in Toronto. A week after that, I trekked off to New Orleans (someone's got to do it). There, I conducted some in-depth cultural research.

Research.

Research.

While in New Orleans, I also attended the Collision marketing conference, where leaders from IBM Watson proclaimed that "deep learning and AI are everywhere." And therein lies the gap between the future promise and current state of AI. With the help of advancements in data and compute power, we have come so far since the term AI was coined 60 years ago, and we have so far to go.

 
ai applications today.jpg
 

Fortunately, AI is at a stage where, with a little help, ordinary folks can still understand the basics. We'll start with some definitions of AI, machine learning, and deep learning, which are often used interchangeably but are different things. Machine learning is a subset of the broader field of AI, which is about creating computing devices that mimic human intelligence. Machine learning is a way to let machines learn using data.

Artificial intelligence is a field within computer science that attempts to design intelligent behaviours in computers. Machine learning is about how to solve AI in a very particular way - using data. Within machine learning, there is deep learning, which is artificial neural networks - borrowing some inspiration from human brains to learn from data.
— Inmar Givoni, Autonomy Engineering Manager at Uber Advanced Technology Group
 
Machine learning diagram from IBM Watson, at Collision conference.

Machine learning diagram from IBM Watson, at Collision conference.

 

Inmar also talked about three subsets of machine learning, namely supervised learning, unsupervised learning, and reinforcement learning. Her example involved using computer vision (i.e., image recognition, like Salesforce's Einstein Vision) to differentiate between images of dogs and cats.

Supervised learning is the simplest. You give the computer lots of examples, and for every example you tell it the correct answer (this is a cat, this is a dog). In unsupervised learning, you don’t tell the computer the answer - you say there are two different things, but you don’t say which is a cat and which is a dog. Then there is reinforcement learning, which is more applicable in robotics or games. The algorithm gets to try different things, and if it gets it right, you say ‘you got it right.’ If it gets it wrong, you say, ‘you got it wrong.’ So reinforcement learning is in between.
— Inmar Givoni, Autonomy Engineering Manager at Uber Advanced Technology Group

Inmar also talked about the value of smaller firms' leveraging AI tools and platforms developed by tech giants, since they have the access to massive data and resources to build these platforms (think Amazon's Tensor Flow or IBM's Watson Studio). At Collision, IBM's Shantenu Agarwal did a live demo of Watson Studio, showing off the ease of use of the platform's pre-built models.

 
 

In terms of applications, most AI today is used to classify ("cat or dog," "working or not working"), predict ("you might like" recommendation systems), detect (fraud detection via pattern recognition), or translate (speech recognition, which I think is pretty cool).

The types of problems you can solve with AI today are pretty narrow. Anything we can define with examples and labels is a good example of what we can use AI for. Like translation systems, or speech recognition, because you can say ‘this is the text translation of this audio file.
— Inmar Givoni, Autonomy Engineering Manager at Uber Advanced Technology Group

These translation technologies exist today. For example, with Skype Translator, you can speak in English, and the person on the other end can hear what you said in Spanish. And speech recognition is employed by a variety of AI-enabled chatbots today.

AI is also being used to solve problems involving probability. Steve Irvine, founder and CEO of Integrate.ai, explains that much of the benefit of AI is in using data to tackle problems that do not have completely predictable, black and white answers - but are instead based on predicting the most likely outcomes. We are moving from a world of certainty to a world of probability, and AI is enabling that.

One of the biggest probabilistic problems is in self-driving cars - because pedestrians and other drivers do not always behave the same way in a given situation. Givoni explains that much of Uber's work is centred on modifying algorithms that improve perception (ability for the car to identify objects in its field of view), prediction (a probabilistic assessment of which way the objects, including pedestrians, are likely to move in a given situation), and motion planning (which way the car will move given its prediction of the way the other objects will move).

I have always loved uncertainty, because in times of great uncertainty, she who can use data (and courage) to make a decision anyway will be miles ahead of those still frozen in fear. The great promise of AI is that it will use the power of data and computing to help humans make better, faster decisions based on assessment of probabilities, pattern recognition, perception, and prediction. And that's a damn good start, even if the Singularity's a long way off.

Meanwhile, in the middle of the conference, the Collision organizers announced that Collision 2019 would be held in...Toronto. This prompted several people at the conference to ask me whether Toronto the Good was as much fun as New Orleans. Ummm...laissez les bons temps rouler. See you next year, fellas.

 
Lobby bar, Royal Sonesta hotel, 11:30 pm, Wednesday, May 2, 2018

Lobby bar, Royal Sonesta hotel, 11:30 pm, Wednesday, May 2, 2018