Es Lee’s first efforts to bring emotional intelligence to texting didn’t go quite as planned.
He’d helped discern the romantic intent behind some texts a friend had received and figured an algorithm could more objectively detect meanings often lost on people.
So he whipped up an app called Crushh that told users how much the people they were texting really liked them. However, the algorithm’s brutal honesty proved a tough sell: who wants an app that often makes people feel bad about themselves?
“I’d like to publicly apologize for Crushh,” quips Lee, an expert in quantitative modeling. “It turns out that if you’re in a relationship where you need an algorithm to tell you if someone likes you, you’re probably not going to find out something you’ll like.”
But like most good entrepreneurs, Lee simply looked at Crushh as a first iteration and got back to work. He’d learned from it that using algorithms to analyze texts worked. It just needed to be expanded to read relationships generally, not only amorous feelings.
The result is Mei (pronounced “may”), a messaging app with an AI assistant designed to help improve relationships. On top of giving advice and observations from a text conversation, the assistant can build personality profiles of users based on their texting histories and add to them with every text.
By learning from how often two people text and what they say, Mei’s algorithm can build a pretty complete picture of how they feel about each other, says Lee, founder and CEO of the New York-based startup.
Mei can start to understand how a relationship is progressing and, at its most powerful, illustrate to users that their assumptions may be off base.
“Just because your ex responds with emotionless, three-word responses doesn’t mean that’s how he or she feels,” said Lee.
The initial obstacle to building Mei was a complete lack of data with which to train a deep learning model, Lee says. But as more people use the app, the more anonymized data Mei has to continuously train its models.
Mei currently boasts 25,000 active daily Android users and an iOS version is expected early next year.
To get personality profiles, the algorithm applies the five-factor model of personality traits to entire texting conversations. Inference happens in real time, with the app sending texts to one of Mei’s two machines running NVIDIA TITAN V GPUs. The company’s growing pool of data is stored on an Amazon Web Services database instance.
Mei was also designed with privacy in mind — all settings are defaulted to “off,” making everything an opt-in choice. All message data is encrypted except when being analyzed on the GPUs. The company doesn’t collect names or any other identifying information other than phone numbers, which are hashed.
Overcoming Limited Resources
As a small startup, Mei is looking to partner with researchers to offer access to that voluminous and anonymous data in exchange for help crunching it. So far, they’ve partnered with a small number of universities.
“They have infinite resources and not enough data,” said Lee, “and we have tons of data and not enough resources.”
Putting that data to work for users in as many ways as possible is key for Mei to reach its potential, Lee says. The company hopes that by sharing anonymized data with researchers, society could begin to learn from the mountains of data being generated these days.
The company is working on algorithms for everything from matchmaking and mental health monitoring (think detection of mood swings) to micro-targeted polling (such as asking 100 doctors if the mole you have warrants having it diagnosed).
What’s more, Lee expects to expand Mei’s reach into users’ communication tools as laws such as the Europe Union’s General Data Protection Regulation increasingly shift control of data to consumers.
Once Mei can analyze messages from apps such as WhatsApp or Facebook, users could get a more holistic picture of their relationships, and perhaps their health.
The post Perplexed by Texts: AI Providing Insights into What Messages Really Say appeared first on The Official NVIDIA Blog.