AI Luminary Series: Reza Zadah Explains the Difference Between AI and Machine Learning
Artificial Intelligence Luminary: Reza Zadah, CEO and Founder of Matroid
In this video, Reza Zadah clarifies the difference between artificial intelligence and machine learning, and what role algorithms play in these fields.
Read the transcript:
The working definition of artificial intelligence now has become maybe 50 different definitions…. It is many different tasks that humans ...are good at but computers are not.
For a long time we were trying to replicate our thought process by putting in a lot of different rules into the computer, by programming them… a lot of logical rules that went one by one — and the computer could follow them, and eventually we thought if we had enough of these rules, we could come up with AI.
That turned out to be a terrible way forward, in that many of the tasks that we can now solve cannot be solved using that approach of write down a bunch of rules. Instead, we write down an algorithm that can look at a lot of data and learn from the data.
Machine learning… is this idea of marrying algorithms and statistics, learning from data. Deep learning is a subset of machine learning. So we build these algorithms that have lots and lots of numbers that we don't know how to set, and then we set those numbers by looking at data. That's the general machine learning task.
It's like you have this algorithm, you have this program that has many, many, many knobs, and we don't know what these knobs should be set at, so we just set them using data, and we call that learning.
That is a very powerful technique that has gotten us very far, and deep learning is a subset of it. The innovation in deep learning… has been setting these knobs using more and more complicated models.
Matroid is a computer vision company. We focus on the long tail of objects that might be interesting to a human in media. So if you have a eight hour video and — and you're interested in finding particular events that happen in this video without having to watch the whole thing, you come to Matroid and you build a detector for it, and then you look for those things that you're interested in, in a few minutes as opposed to eight hours.
For these very complicated models, we can't — we can't just have a simple rule that tells us how to take the derivative of these models. Instead, we use a computer to compute the derivative. Once we have a derivative, we can use that derivative to learn from data.
With regard to what the community can do to help AI along, There are many ingredients to machine learning. There's the algorithm, there is the data that — that the algorithm is learned from, and then you also need to learn how to run these algorithms and how to build them in the first place. So for the community to move forward, we have to make all of these things available.
I'm looking for that problem where if you can solve this problem, you can almost certainly replicate a human. And that problem is out there still. We don’t know it.
To really try and make progress, we need to define the tasks that are — that are most crucial to replicating our thoughts.
So, provide the motivation, provide some baseline ideas, provide the monetary motivation as well, and that will spur a lot of research. That's, I think, the future of AI.