Boston-based AI startup Gamalon is launching two new products based on the innovative technology, MIT’s Technology Review reported today. The company has developed a technique that gives artificial intelligence the ability to write code. It uses a predictive model that’s based on probability to work out facts relating to data it’s supplied with.
The information it gleans from the first stage is then used to create a more optimised version of the model. This powers future runs of the program, enabling more facts to be determined the next time around. The result is an efficient and effective self-learning procedure that lets the AI develop new skills on-the-fly.
In practice, a first run of the algorithm could determine that cats are likely to have ears, whiskers and tales. By studying further examples, the AI could then work out more about cats, without being provided direct data about the animals.
In essence, the system is built on a chain of algorithms. Each algorithm requires less data than the one before it. It is capable of creating a successor that’s even more succinct.
Gamalon has already applied its technology to a drawing app that can work out what the user’s trying to sketch. In much the same way as a spectator gradually realises what you’re drawing, the app comes to recognise meaning in the way you position objects. Adding two dots and a curved line to a circle will make it think of a face. A triangle on top of a square could be a house.
The technique has been studied before but this is its first mainstream use. While potentially very powerful, it’s also extremely demanding. Traditionally, it has required multiple processors to run in parallel for days. The difficulty comes from having to consider so many possibilities and then determine the most probable result.
In the example above, the cat could have ears, whiskers and a tale, any two of the three or just a single one. Even this simplistic scenario has at least seven potential outcomes.
Despite the challenges of working with such a complex technique, Gamalon is optimistic that probabilistic programming will have a significant impact on next-generation consumer tech. Automated vehicles could use the system to recognise new obstacles as they approach, giving them more scope to respond safely. On the desktop, next-generation digital assistants could work out your interests in minutes without requiring training, making technology work better for you.
For now, the technology remains limited to commercial cloud applications, requiring hefty processing power to run efficiently. That’s not to say there’s no products using it though – Gamalon already offers Structure, capable of extracting raw text from product descriptions, and Match, a utility for categorising products in a retail inventory.
