<img height="1" width="1" style="display:none;" alt="" src="https://dc.ads.linkedin.com/collect/?pid=193356&amp;fmt=gif">
Demo
Login
Menu
Demo
Login

[VIDEO] What real business problems can be solved with AI?

by Tyler Foxworthy, on March 15, 2018

What is the potential of AI? Tyler Foxworthy, Chief Scientist of DemandJump, sifts through the noise and talks about the real value or artificial intelligence now and into the future.

artificial-intelligence-go-beyond-the-hype-webinar

Transcript

Tyler: Hi my name is Tyler Foxworthy and I'm the Chief Scientist at DemandJump. I lead our research and development efforts.

I get asked a lot about AI.

What is AI, and, probably more importantly, what types of problems can AI solve, and how does that relate to me?

There are wide variety of businesses problems that can be solved with AI, even outside of marketing, which as you know is where we live and breathe.

I like to think of it as falling into a number of primary buckets that are really application specific, so I'm just going to roll through a couple of these major areas and then talk a little bit about applications and use cases.

So the first big bucket for machine learning (or AI or whatever we want to call it) is optimization.

Optimization means, in the way that I'm interpreting it in this context, is how do we use data and information about processes in our business to in order to either minimize risk, minimize costs, maximize gains, and do that is very systematic.

A great analogy that I like to give - and this is also a great application of algorithmic optimization - is portfolio optimization in finance.

So when a mutual fund is being assembled or an ETF is being assembled, managers are often algorithmically looking at large numbers of individual securities and they use mathematics to figure out which combinations of securities, when coupled together into a portfolio, will minimize the overall risk volatility while maximizing expected gains.

We're also starting to see applications here beyond just traditional finance. If we look at the historical record in industry, Six Sigma was the precursor to a lot of the modern algorithmic optimization techniques that we're seeing in advanced manufacturing today.

Andrew Ng, the former Chief Scientists at Google, Baidu, and Coursera (very prolific) has actually just recently launched a new fund and company devoted to bringing AI and optimization techniques into traditional manufacturing and less sexy less sexy industries.

I think he's right. I think that long tail of business is currently being very underserved by by our modern technology and that there's a tremendous amount of gains to be had there, both large and even in small to medium-sized operations.

Another huge bucket that that we think about a lot is forecasting. Think revenue forecasting, sales forecasting, demand forecasting.

You know, these techniques traditionally were the purview of classical statistics where we would use techniques like moving averages, ARIMA, and detecting seasonality in order to help understand where these latent trends and data are. Then we would try to extrapolate some future event.

Over the last 10 - 20 years, we've seen an explosion in the development of techniques they're using that are relying on things that we would call more machine learning or advanced inference Bayesian optimization in order to make far more accurate and far more flexible models for doing forecasting.

Another huge bucket is anomaly detection and we should all be familiar with this in our daily life with credit card fraud, for instance. How often does one go out of town and get their credit cards cut off because some algorithm somewhere decides that you normally spend within a certain radius and you're outside your buckets. The algorithm learns where your map is and if you go outside of that "learned bucket" it's going to raise a red flag in order to say, "Hey, this is normal or this is not normal and that's what anomaly detection is.

Anomaly detection isn't as distinct from classification and that's going back to the idea of AI or machine learning as a yardstick. What we're really doing is taking sets of patterns and sets of observations and we're measuring them. Then we're measuring the distances between pairs of observations and so an anomaly is just an observation that, when we compare it to every other observation, is past some critical distance beyond which we say we want to know about this.

Anomaly detection has a tremendous number of applications both in finance and things like credit card fraud and tax fraud. I have personal experience in this area and in developing algorithms actually for tax fraud detection based upon these principles.

Another huge area of application is automation. This is actually where I began my career working in biotech developing algorithms to automate the diagnosis of various disease states based upon blood and urine and the chemical analysis and genetic analysis of samples to determine whether or not this observation was was normal or not normal. It was a highly automated process that used to need somebody with a master's degree or higher to interpret these results. We developed algorithms that were able to do all that work on its own with minimal human intervention which is a huge value proposition for for AI in general.

Comments

DemandJump Blog

From display & search marketing wisdom to general growth tips, here are a few resources to help your digital marketing teams work smarter, not harder.

Subscribe to Updates

Customer Experience Update