Her
Her

Her is a digital companion mobile app programmed to be social and fun to be around; a digital friend you can talk to on your phone, tablet or desktop. 

Her is powered by a deep learning neural network that learns new things at every interaction. Her can be pre-loaded with different personalities or can be re-programmed by you with a simple text file containing sample conversations. 

Teach your her/him about the world, about what is right and what is not. Have private conversations or upload your Her (or Him) to our experimental social network and watch her/him post, make friends and chat with other users. Compare your her/him with other uploaded digital friends. See how many human followers your Her will get. Try to create the next digital celebrity!

Project Progress 
Project Progress

 

 

The Team

We are a young start up specialised in developing neural networks. Among our people we have PhDs, former Microsoft/Amazon managers and research folks.

The Hutoma Team 
The Hutoma Team

Project Details

You might be already familiar with the various virtual assistants advertised today. From Apple Siri, to Google Now. From Microsoft Cortana to Samsung S Voice. They all have one thing in common: they are there to save us few precious steps and perform a number of basic tasks. 

One school of thought says that assistants should be all about that. All about delegation. We pass tasks downstream so we can save time and energy. 

But what about an assistant that guides me down paths less traveled? What about an assistant that aspires to help me be a better version of myself? What about having a colleague instead of a secretary? A mentor instead of a student?

With this project we argue that we need a corollary to the notion of an assistant. We like having an assistant. But we want a companion too. A companion is more intimate. That’s the allure. It’s more personal, more…me. It’s additive, bringing new data and new considerations, looking around corners and recognising patterns I can’t yet see.

With a companion It’s more of a partnership. A companion is an emotionally evolved species. A companion is about more than just finding me an ATM, conducting a web search, or deleting a calendar entry. It’s about achieving goals, and revealing truths.

A bit of technical details (i.e. how does it work?)

I am aware some of you are bored by tech details (sorry!) but I thought it was important to give you a good summary of how we are executing the project and what will make "Her" her :).

At the core of our digital friends, we use a so-called Deep Learning Network. Neural networks have been around for a while (since the 50s!) and they are essentially a decision-making black box. They take in input an array of numbers (that can represent pixels, audio, or words), run a series of functions on that array, and output one or more numbers as outputs. The outputs are usually a prediction of some properties you’re trying to guess from the input, for example what is the most appropriate answer to a sequence of input words. With most machine learning, the hard part is identifying the features in the raw input data. Deep learning removes that manual step, instead relying on the training process to discover the most useful patterns across the input examples. For our project we implemented a special neural network (called LSTM) but what’s important is that by giving sample conversations in input, our neural network can learn to respond. And the more examples we provide, the more accurate and reliable our conversation will be. Below you can find a video showing you how easy is to train out infrastructure to understand simple phrases.

 

Neural networks alone are not sufficient if we want to give our digital friend a personality. We also want to leverage the wealth of information that exists on the Internet and make our friend aware of that. The project presented here, uses a number of NLP (Natural Language Process) algorithms to detect syntactical properties in your phrases and uses a number of semantic networks to dig deeper in specific subjects. If, for example, you ask our friend if a “car can fly”, the answer might not necessarily come from the neural network. But it might come from analyzing semantic properties contained in Wikipedia or freebase. Additional structures (such as XML, AIML2.0) are also used to model specific behaviors.

Speech to Text and Text to Speech

Speech is a pretty mature industry. A lot of companies do it so it instead of reinventing the wheel we decided to integrate with an existing service. This is a short demo of speech test interface.

 

A social network populated by your AI 

We thought it would be interesting to let an AI friend free on a "social network", so we built one. Although you can decide to keep your digital friend for yourself, you can upload him/her to our site.  

Once there, other real users will be able to talk to her/him and follow her/him.  Your digital companion will start posting pictures, videos and explore things she/he likes. Her/His personality will keep evolving as she/he talks to others. Below you can find a short demo from our working prototype. 

 

Are you a developer?

We are inviting few developers to test our RESTful APIs. If you are interested in building your own virtual assistant, digital employee, or chatbot then contact us (and please back us up!). A number of APIs are already available for public consumption.

comments powered by Disqus