Interview with Jordi Torras of Inbenta – Using AI to Improve Website Search

inbenta-AI-pyramid

Voicebot had the opportunity to hear Jordi Torras speak at the RE-WORK Virtual Assistant Summit in January and we circled back with him recently to ask a few more questions. Torras is the founder and CEO of Inbenta, an AI-based search company that helps automate and improve customer service and eCommerce functions for a company’s customers. He wants to “give users the power to find relevant information online using natural language technology.” Torras relocated from Barcelona to San Francisco to make it all work. Sounds like just the right background for a Voicebot interview.  

Looking at your background, it appears AI is not new to you given that you were working with expert systems in the 1980’s. Tell me about what the difference is between your earlier work in expert systems, why they didn’t work out and how things have changed.

Jordi Torras: In the 1980s there were many different techniques. It was an explosion of alternatives. There were expert systems, neural networks, inference engines. It was a time of Prolog, of LISP. These were programming languages for artificial intelligence. There were even very cool new things called genetic algorithms. They simulated populations of solutions to fit for one specific problem.

But, none of these really worked. The expert systems didn’t work because it assumed you could extract rules from experts. The experts will tell you a number of rules, you put those rules into a system and the system would be as expert as the human. But humans don’t think that way. It was obvious that humans might have rules, but it is hard for them to show you what rules they were using. Credit scoring systems were okay. There was one system that was able to diagnose infectious diseases based on rules. After a number of questions the system could suggest a diagnosis. But most of the expert systems were finally abandoned because they didn’t work very well.

The reason why neural networks didn’t work back then is because we didn’t have good algorithms. Everyone understands that neural networks are designed to mimic the human brain. However there are more neurons in a brain than stars in the galaxy. You can’t replicate that in a computer. Today you can mimic the type of brain in a fly. And that seems simple, but flies can actually do a lot of things. Fly. Eat. Avoid obstacles.  

The equivalence between bio neural networks and artificial neural networks [is based on the connections]. It is a model copied from biology, but we have to stop [at the structure] because we don’t know how we learn. After that, it is all math. We have mathematical models to make learning happen.

Deep learning is a mathematical invention to translate data into knowledge. Human neural networks are pre-programmed knowing many things. Since we don’t know how we learn, AI is using math. In 2015 it started really working on deep learning. The principal was simple. The new system would teach every layer of neurons to solve a specific task. They started with determining if a picture contained a cat. The first layer was looking at shadows. The second layer was identifying eyes and noses. The third was animal forms. They were doing something interesting with neural networks – classification. Instead of experts telling us the rules, the system would learn the rules by itself by reviewing examples.

On the one hand you need a lot of data, but we have a lot of data. All we needed to do was label those images. That system has exploded the AI.  This is why AI is back. On the one hand computers are faster, on the other there is so much data everywhere that there is data to train and deep learning can infer rules that even humans cannot. That is the difference.

We now even have AI reviewing medical imaging. It is very hard to find small differences in images. Humans are good at it and the AI is learning. However, the AI can see metadata that is collected by the imaging system but isn’t made available to the doctors. You can’t see it in the image, so the doctors can’t view it, but the AI has all of the metadata and the data that can be seen.

What is surprising is a computer model that was copied from a biological model worked that well once we understood how the neurons worked. But we didn’t know how they learn [and still don’t]. New algorithms will improve existing algorithms and AI will grow. For many years it was all about the CPU and memory. Now it is about simulating artificial neurons very quickly. Human neurons are actually slow. We are slow computers compared to real computers. What we have is parallelism. For certain things computers will react faster than any human can react. We can’t use the brakes in a car as fast as a computer can.

What led you to form Inbenta?

Torras: When it comes to search, there is the internet and that basically equals Google. Google has 99% of the market for internet search and it works pretty well. Then there is this other search within websites. These tools suck. Everyone is frustrated. Why is Google so good searching the internet, but search engines within a website are so bad? Google launched search appliance so people could use Google for their website. Then you had Google searching their documents. But, it still sucked. Why is Google working on the internet and not in the enterprise?

There are so many places where this is important. Customer service. eCommerce. I founded the company to make search within a company great. That is the original vision. We believed that by using computational linguistics, AI and a theoretical framework called meaning text theory, we could go to the next level. The system would understand what you were trying to search for and it would deliver the right results.

Inbenta users want the solution to deliver an answer that is relevant. We ask customers to have a good list of FAQs. When you ask your question, it provides the right answer back. Often questions and answers don’t have any common keywords. That is what Google cannot solve today for the enterprise. Inbenta has helped some companies reduce their call center staff by hundreds of people.

I understand that you started with a goal of replacing keyword search. How has that idea evolved and taken you to where you are today?

Torras: We have evolved. We also use machine learning and neural networks and deep learning. We combine our engine with the latest improvements in AI. The way it works nowadays is all of our software is delivered through the cloud. We use AWS. When machine learning broke in, the hosting services were very quick to provide these open source algorithms and integrate them into their cloud services. So we access them as a service. It looks like AI is part of the operating system right now and it happened very quickly.

The algorithms are less important than how you prepare the data to learn and how you use the output from machine learning to improve the system. It is very easy to go into machine learning and have a system that doesn’t learn. There could be not enough or too much data. There could be too many neurons in your neural network. There are a lot of details that can make an implementation work. What we are using now is deep learning algorithms that you buy as a service. There are open source libraries but it is easier to go to the cloud and get it there.

As a side-note, tell me about the automated long-tail SEO solution.

Torras: Many customers go to Google to get answers instead of a company website because it works better than the search on the actual website. It is the Googlization of customer support. Your customers might now find your answer because they will go to Google first. We publish landing pages for specific, successful and useful questions [that the Inbenta system answers. It puts key words in a page so Google can find it].

You raised a $12mm B round in September 2016. What changed about the business that led to the investment infusion and how are you using the funds?

Torras: We are using the funding for hiring, adding new members of the management team and more professional structure. Also, R&D, product management, professional services, customer success and account managers. We had one person in marketing; now it’s a department of six people. We are creating the structure for growth. VC funding is a time machine. With the money you can build the organization you would have had in five years. With VC funding you can have it today so you can scale up and go faster.

Something must have happened because we entered into the radar of investors. Eventually the momentum and attention AI has received put the attention of investors on AI. The investors found us. We weren’t looking for them.

It seems that most voice and chatbot applications are using AI or neural nets solely for the natural language processing / natural language understanding (NLP/NLU) and then handing off to a rules-based system to return the results. Is there no need for neural nets to pull in or discover more meaningful data for query response?

Torras: If you look at deep learning neural nets, et cetera, they are pattern recognition systems but they do not generate things. They can identify a cat, but they cannot draw a cat. They cannot create. Machine learning today is mute. There is a whole area of natural language generation that not only answers with pre-existing text, but is speaking new things all of the time. However, no one has done that successfully. They are all lousy.

There is a pyramid of the AI. The base is just computing. The calculator computes things very quickly that would take you hours. The next level is automation, ATM machines, things that automate tasks. The next level is narrow intelligence. Then there is general intelligence, as smart as a human. We are not close to that. If we reach that, we will quickly move into super-intelligence.

There are some theories circulating that the rise of AI and voice assistants is similar to mobile, others say it is more like the rise of the web, and still another school suggests it is different entirely from previous technology shifts. What do you think?

Torras: This is something entirely different. In a few months or years a new AI winter will come. AI and machine learning will not be cool anymore. The technology won’t stop, but it will stop having all the media attention. People will understand the actual progress that is being made, but it is not what they expect from science fiction.

What is your favorite voice application, skill or action?

Torras: I use Siri. Sometimes Siri jumps into conversations when I’m not expecting it. The most basic application is my favorite, speech recognition, so I don’t have to type on the little keyboard.

Voicebot Interview with Todd Mozer of Sensory

Voicebot Interviews Arte Merritt, Co-Founder and CEO of Dashbot.io