Thought Leaders in Big Data: Jim Rebesco, CEO of Striveworks (Part 2)

Jim Rebesco: Another challenge that we face is the data-centric AI movement. The data-centric AI movement says, “We can do a lot better in preparing and building analytics if we spend that marginal hour not on playing with model architecture but rather if we think, very thoughtfully, about the data we’re using to train the models.”

This is a trend that is mirrored in the more commercial side of analytics. What does that mean? This means that these data-enabled companies are now seeing this trend of increasing personalization. Maybe the model that works best for me may be very different for you.

Now you’ve got this very focused challenge of how do I know what the right model is. How do I validate the output? Most importantly, how do I audit it? What do I mean by audit? Unlike the law of nature or physics, data science models are statistical models. If I find an example that demonstrably shows that F = MA is not true, it tells me that we need to tear up that natural law.

If I have a data science model or regression model and I say, “This is an outlier.” It’s not a problem with the model. The way we talk about models is very artificial. We talk about good or bad models. We talk about trust or untrusted models. It’s very binary. The reality is that the best models in the world still have finite error rates.

That’s true in things that don’t matter. If I serve the wrong ad on top of a web page, it’s not a big deal. If I do the same thing in prescribing a drug to you and making a diagnosis, that’s an incredibly big deal. What we need to do in these environments is not just build very good models tailored towards individuals. Importantly, we have to have an audit function. We have to be able to say, “Hey, this model made a mistake. Why did it happen? Where was that error used elsewhere in the system? How can we communicate that error very quickly?” That’s one of the key recurring stories we hear from customers over and over again.

It’s understood that models make mistakes, but what I don’t understand is when those mistakes enter my own information processing system, I don’t know how to trace them and understand the downstream consequences. That, ultimately, has been one of the most limiting factors in the uptake of AI/ML in highly-regulated industries.

Sramana Mitra: Let’s take your outlier example. How do you find what’s causing problems? How do you report and alert the model makers what to do with it?

Jim Rebesco: The way I’d say that you should approach that is at three different levels. First of all, how do you know that there is an error? There is no single way of doing it. We typically break that down into three major checkpoints. First is input. Garbage in, garbage out is a mantra that applies to all of us in the data processing world.

For data science models, that can be refined in a more actionable statement. Is the data that is coming to my model in the sample or out of sample versus the data that it was trained on? We know that the data science model tends to generalize poorly. The first yellow flag that I should wave is to say, “Whatever manifold that my training data sits on, this data point is off the map.”

Once you put it through the model, the other recurring heuristic is that humans are pretty good at defining matrix bases. If I give you two different pictures of pandas, those are both pandas. Consequently, if I move a few pixels around, you shouldn’t be very stable in your prediction. You don’t necessarily see that in analytic models. In particular, you see a very nice correlation that as you perturb the input slightly, models that are poorly suited tend to show more variance. What is my robustness and stability measure if I just inject random noise into the input?

I’m a Bayesian at heart. That’s my prior. If I’m getting a yellow flag in this fundamentally unsupervised way, I know this is a suspect data point. From there, it’s driven by the use case and UX that you need to support within the application. If I’m doing healthcare, maybe the answer is to send this data point to a doctor. Ultimately, you want humans to be arbiters, but you don’t want to be dependent on them to do the routine stuff.

Sramana Mitra: I was asking about the level of sensitivity at which you start manual intervention. It sounds like that is specific to the vertical and use case. If it’s a medical vertical, it’s probably a higher standard. If it’s advertising, it doesn’t matter. You can do with a less strict fit.

Jim Rebesco: Right.

This segment is part 2 in the series : Thought Leaders in Big Data: Jim Rebesco, CEO of Striveworks
1 2 3