Thought Leaders in Artificial Intelligence: Ashutosh Garg, CEO of Eightfold.ai (Part 4)

Sramana Mitra: You explained to me the data structure. Can you talk about how the learning algorithm learns?

Ashutosh Garg: That’s our secret sauce. We use recurrent neural network to model some of these things.

Sramana Mitra: How much of this is pattern matching, statistical modeling kind of inferences versus expert system kind of inferences?

Ashutosh Garg: Most of these are the statistical model inferences. We are already supporting template languages.

Sramana MItra: That actually brings me to a trend question. This has been my observation now for 20 years. I am still a little bit shocked that this is where we are. How is it that we are still not enhancing statistical models with more expert system kind of knowledge?

There are best practices. In every workflow and in every industry vertical, there are industry-specific best practices that can be coded in with expert system kind of insights that the learning models can draw from. How is it that that’s not really standard in the algorithms that are in vogue right now?

Ashutosh Garg: A few things on that one. This was late 90’s. You can call it expert systems, but another part over there was model-based learning, which is your parametric models versus non-parametric models in some sense.

With parametric models, you were able to capture human knowledge about the system. You had non-parametrics systems that didn’t allow human learning, or very minimal. In general, you are able to do parametric models with vast amounts of data, but the challenge was that if a model is not correct, you will never achieve optimal answers.

Over the last 15 to 20 years, what has happened is that the amount of data that is available to you for learning has increased meaningfully. Computational power is also available. The good news is that now we can have non-parametric models that can learn more easily. The bad news is that people are not even thinking about what they’re building and modeling.

Today, you go over to one course in Coursera and you’re already a data scientist. Really, what you know is how to throw in the data and launching a model. If you don’t know the model and how you are thinking about it, you get into all kinds of biases. We have seen that happen.

The positive side of that, though is, if you don’t know what non-parametric model to use or how you should think about it, you can mess up things. If you are building an expert system, you’re also relying a lot on human intelligence. That can manifest human biases which is very prevalent in our workplace. In recruitment, you see a lot of biases against females, biases against old people, minority groups.

This segment is part 4 in the series : Thought Leaders in Artificial Intelligence: Ashutosh Garg, CEO of Eightfold.ai
1 2 3 4 5 6