Expert systems and machine learning are two ends of a spectrum working to solve similar problems quite differently. One one hand you have if-then scenarios and a logical approach, and on the other you have vast neural networks and a big data approach. Some companies exist to try and bridge the gap between the if-then rule systems and the massive piles of data. They hope to find a middle ground of sorts, one that mitigates their individual disadvantages. One such company is Montreal’s fuzzy.ai.

In this episode, we interview its founder, Evan Prodromou about the state of the middle ground, so-called hybrid systems. The middle ground is an elusive, still mostly theoretical concept, but businesses can take steps to prepare for when it becomes accessible to them. What exactly would a hybrid system provide to businesses in terms of automation? How accessible are they now, and what can businesses do to best integrate them when they’re ready? Find out in this episode of the podcast.

itunes-podcast
itunes-podcast
stitcher-podcast
google-podcast

Expertise: Expert systems, machine learning, small business applications of AI

Brief Recognition: Evan Prodromou is the founder of fuzzy.ai, a company which claims to allow businesses to build AI systems for price optimization, product recommendations, and lead scoring without the need for networks of data. He is also the founder of social network sharing software company E14N, the current chairman of the Social Working Group, and a director at Open Cloud Initiative. He holds a BA in physics from the University of California, Berkeley.

Current Affiliations: fuzzy.ai, E14N, Social Working Group, Open Cloud Initiative

Readers and listeners with an overt interest in developments in core machine learning technology may want to listen to our interview with Dr. Hanie Sedghi at the Allen Institute of Artificial Intelligence.

Dan Faggella and Matt Fogel of Fuzzy.ai

Me with Fuzzy.ai’s other Founder, Matt Fogel (on the left) – at Larry’s coffee shop in Montreal, during my 10-day AI interview trip to Canada in July 2017

Big Ideas:

Evan suggests that some smaller businesses could benefit from making use of the middle ground between expert systems and huge machine learning networks. Not everyone has a large dataset available to them with which they can create a machine learning system. Small businesses can, however, use the knowledge they already have about their customers or their domain in order to come up with hypotheses they can test using a hybrid system that makes use of both logic and data.

Evan uses dynamic pricing as an example. How would a logic-AI hybrid system determine which price a visitor sees when they visit a sales page? It’s common knowledge that customers are generally more likely to buy when prices are lower (increased conversion rates). Based on that understanding, Evan suggests encoding a rule in the if-then system in which the system checks if the visitor has purchased anything from you before. If they haven’t, they could end up seeing a cheaper price in order to make it more likely that they’d purchase – which would allow the company to retain them in the long term, making further profit from future purchases. An existing customer (one who has already purchased) might not get such a discount because trust with them has already been established, and their email details (for future promotions) has already been collected.

If the product on page is a popular one, Evan explains the system could run another check to determine how well the product is selling. If it’s selling well, the system might actually determine that lowering the price for a potential first-time customer isn’t necessary because he or she is likely to purchase anyway.

Even also suggests that the transparency of a system like this is useful for small businesses because they won’t require IT experts to work within it and they can still garner useful insights for their business without potentially expensive, complex machine learning systems.

Turning Insight Into Action

Business owners might consider jotting down some of the things they already know about their customers or their domain.

  • Can you do a quick search and figure out if your customers tend to purchase more frequently at a certain time of day or on certain days of the week?
  • Do you make more sales when you use more images on your sales pages or in your emails?

Business owners can turn the answers to questions like these into rules. When they first implement a hybrid if-then system, they might learn a lot of what they already know, but they can then tweak the rules they’ve already created to see if anything changes. If anything, a hybrid system is a tool to conduct experiments and collect data. Maybe one day, with enough of it, business owners can decide to move onto a more robust machine learning system.

The middle-ground, however, is something that still requires a lot of research. Evan suggests three actions businesses can take in the interim in order to prepare for it when it becomes accessible:

  1. Look into having multiple different deployment options for your business: open source systems, cloud sources, licensed software.
  2. Don’t look to create AI systems that are too opaque and that require experts to work and understand. Ideally, someone should be able to learn how to use your system in an afternoon’s time.
  3. Start using AI in small places to get used to it. Make AI integration a part of your daily business habit.

Interview Highlights on the Evolution of ML with Limited Data

Dan (3:30): How would logic work with regards to dynamic pricing?

EP: First, not all of us have a million transactions records lying around. It really is the problem. Even if you do have millions of transactions records available, they might not be applicable for the product, and you need to have a lot of data in order to get intelligent machines out of those systems. Second is that In order to do that, you need to have some pretty widely varying prices in order to get some kind of good statistical coverage of your possible pricing range.

The last thing is people are spending a lot of data capital to get some pretty basic insights into their system. What we do with hybrid logic learning systems is we start off with things we already know. Customers prefer lower prices. We know that, and we can encode that in a rules language. We can encode things like “first time customers, you want to give a big discount to in order to get them to become long term repeat users.”

Popular products you really don’t need to give a big discount on because they’re obviously selling well. You can put together very straight forward rules based pricing algorithm with 10, maybe 20, rules that can be very simple and really reflect basic business information

Dan (6:25): Could these 10 or 20 rules apply across an entire site or just particular products?

EP: You can do either a ruleset that works across your whole site or you may have a more specific ruleset because you know more about the customer. Generally what we do is a single pricing algorithm across an entire site. Once you got that set up, when a user lands on your site, you can in real time do the calculation for what price you should show them for that product.

If they’re looking for a pair of shoes, you can calculate for how long they’ve been registered, how well the product has been selling, day of the month, time of day. Then, we’re at a moment where we generate a feedback event: either the user buys the shoes or they go do something else. Either way, we can take that information and use it in order to optimize the decision making. Whether we made $225 dollars or we made $0, we can try and optimize the system more.

It’s a very simple mechanism: first you build your rules, then you have inputs that come into those rules, you get an output system, and then you use feedback in order to optimize the decision making.

Sometimes the changes are small. We have an idea that people buy at 2:00 in the morning, and it turns out it’s more like from 1 to 3. The optimization process might be taking some of those hard coded parameters and making them fuzzier. Another thing is the relative importance of rules. You may think the most important thing is the age of the customer, but it turns out it’s their level of income or their location. These optimization techniques are relatively well known. They’re what make machine learning work.

The really nice part about this hybrid system is that, first of all, it’s not opaque. You can just look at a decision and say, “clearly the system was looking at the user’s age, and this is what came out, and this is why we made a decision.” In a machine learning model, that could be buried deep in a neural network, and so it’s hard to audit where the decisions happen.

The other thing is that it’s very robust with regards to making changes or adding or removing data. It may turn out you factor gender into your decisions and that may not be a factor at all, and so the optimization would phase out that parameter. It’s a technique that can add or remove kinds of inputs. It’s very transparent for developers.

Dan (13:33): It seems like we’re trying to find equilibrium here. Can we still understand the core premises at work and know what a decision is an why a decision is an at the same time leverage machine learning and have that actual learning happening in real time, not just routing through simple yes/no gate?

EP: I think that AI in terms of business decisions is in its early years. I think that over the next decade we’re gonna be finding different kinds of applications we’ll use different techniques for. There is a wide array of different applications within there, and different companies are going to find what works for them

Dan (15:35): One of the dynamics that your business is predicated on is that there will be a middle ground between unlimited data machine learning and ice cold yes or no systems that can solve a lot of problems. Is that correct?

EP: Today in 2017 we really think of AI as this kind of moonshot business process. That what you’re going to do if you’re a bank is you’re going to have a robot teller that’s gonna do everything for a user.  We got this idea that AI is this one huge thing. I think it’s much more likely that AI is going to be an important part of the software development practice.

We need to start looking at places in those software systems that we can start putting self optimizing intelligent features, whether its showing the price on your ecommerce site or whether, for an internal system, deciding about calendaring or figuring out where the best place to fit a meeting is. Most of us are experimenting with data driven techniques like AB testing. Starting to put some AI behavior behind that is going to be the next step there.

Dan (19:34): What’s it going to take for these technologies to become more accessible to business?

EP: I see three big trends:

  1. First of all is going to be deployment option. If you’re deploying an [inaudible], you have a lot of work to do in order to set up your cloud system and support that open source product. That’s really your only option. I think that we have to have different deployment options for different kinds of companies and different applications. Maybe cloud sources, maybe open source systems, maybe licensed softwares. We have to be able to make the decision that is right for the application, not just the technology.
  2. Second is developer experience. If developers have a hard time getting started or if the system is opaque, you’re dead in the water. We know [at fuzzy.ai] we wanted to have people be able to quickly come in and build intelligent systems.
  3. Third is a change in mindset. Instead of thinking about big projects, think about dozens or hundreds of small places to do AI. That’s the only way you get the institutional expertise and understanding of where it can or can’t work for you. If we can keep the barrier to entry low, then software developers should be able to turn to AI as a daily thing and build AI systems for their daily business.

 

Head image credit: Northwestern University