The interview was originally posted by digamma.ai. 

1. Fido’s technology powers chatbots so that they can learn automatically from the internet and answer complex business questions based on a large volume of text. Tell me about how Fido got started in the first place and your novel approach in teaching machines how to reason more like humans.

In 2003, we co-founded Fido Interactive and implemented many commercial projects of AI systems, including chatbots. However, we slowly started to reach the limits of our capacity of building handcrafted knowledge bases. As a consequence, in 2007, we co-founded an independent AI lab in Europe to tackle this problem. Our focus moved from creating bots to understanding how a chatbot could read and learn by itself. We knew that the incoming trend of statistical learning wouldn’t have been enough as learning without reasoning is only a way to mimic intelligence within very narrow tasks. Instead, we first aimed to teach the computer how language works step-by-step. Then, we taught it to reason based on how people use language to express themselves. This way, we enable it to learn automatically without any data labeling. One of the first systems we created this way was Cerber. It was a public safety system designed to detect any misbehavior of an adult towards kids in online chatrooms.

2. Your company has developed a new type of hybrid AI that understands the English language at a deep and comprehensive level. Unlike other technologies, it does not require training or labeling of data and acts as an interface that allows deep learning to understand text directly, removing the ‘data neediness’ obstacle that AI often presents. Why is this significant?

It’s significant because, for most real-life problems, we simply don’t have labeled data. And we won’t have it.

One of the fields where this problem is particularly significant is healthcare, which is often characterized as a ‘label desert’. Currently, typical business solutions try to understand the topic of a conversation, user sentiment towards a certain product, or — more generally — to sort the information into a set of categories.

The problem is that all of these categories need to be predefined beforehand and the training of a deep learning model requires tons of examples that need to be labeled with these categories. Imagine how many categories you would need to figure out all the reasons why people stop their treatments. And this isn’t the only thing we would like to know, so the possibilities are nearly endless. Even within a single domain.

And now imagine an AI system that knows how language works and how people can express the reasoning behind certain actions or behaviors. Such a system would no longer need to be trained on millions of labeled examples for every single problem. Fido AI can learn and reason by itself and this can scale to the whole Internet.

3. Tell me about several recent and interesting applications of your technology.

Recently, we can observe an epidemic of suicides because of ferocious online behavior — the idea is to educate individuals about the dangers of cyberbullying. We developed a system to detect cyberbullying and determine if any violence or bad behavior is taking place online. This project is a great use case demonstrating an application of the third wave of AI.

Another application of our technology involves figuring out why people switch off of their medications, what’s happening with people after they take the medication and what might be the side-effects or off-label uses of the meds. Such information can help pharmaceutical companies improve their products.

With car manufacturers, we are working on an in-car navigation system which can direct you to the best and safest places to run, a park where you can walk your dog off-leash, a good parking spot and even tell you where is the best place to clear your mind. It all starts with a system that reads and understands travelers’ conversations online.

The key point is that we enable computers to learn by reading, not by simply being fed with handcrafted data or trained on labeled examples. The challenge lies in teaching machines how to successfully learn, read and comprehend. As a company, our own challenge is that our technology is currently language-dependent, meaning that we are bound so far to only one language, which is English.

4. An independent benchmark study compared the accuracy of both Fido and Google language parsers. The conclusion of the report was that Fido’s technology achieved double-digit better accuracy than Google’s AI on human-written reviews and opinions. What is it about Fido’s technology that has contributed to this impressive result?

The reasoning part makes the difference and contributes to the different level of accuracy. Google’s parser is built solely using the second wave of AI, which is statistical learning. We have added reasoning to its already impressive learning capabilities, which — according to DARPA — represents the third wave of AI. It’s worth noticing that the parser is not the AI itself. It allows machines to decode the logic and the context. From there, we are able to harness the context and teach the AI system how people can express themselves — this is the foundation of automatic knowledge acquisition.

5. Fido already knows how people can express their fears, warnings, and recommendations. What’s more, it can understand our needs and find the best places for our unique profiles based on the world’s collective intelligence. Tell me more about this capability.

Fido’s vision is to harness the world’s knowledge, which is hidden within the internet and make it available to everyone who wants to learn from it. The data is out there. The only problem is that currently, we would need to read it all by ourselves, which is nearly impossible for us as individuals.

Google’s goal is to index all information online. We want to give you the answer you’re looking for, not just a link to a website. We want to build a ‘collective intelligence’ —not to mimic a human brain, but rather to get access to the knowledge from billions of humans’ minds combined together.

Comment