Getting computers to write programs

Article By : Rick Merritt

Today’s deep-learning approaches will remain useful for tasks including images and video recognition.

« Previously: Software explores faster route to machine learning
 

New start-up Gamalon has launched two products. Gamalon Structure converts text paragraphs in databases or documents into rows of structured data. Gamalon Match can de-duplicate and create links among the data rows.

The 14-person company is working to bring up multiple customers while it extends its core technology beyond text data. It has produced videos of lab demos of the system recognising objects sketched on a tablet after only a few examples.

__Figure 1:__ *Gamalon claims much lower training needs to recognise a text phrase with high accuracy. (Source: Gamalon)*

Long term, Vigoda believes that his technique will let computers develop their own programmes. They will both create large programs out of smaller modules and interact with human programmers able to read its high-level Python output, he said.

He envisions groups of specialists around the world using Gamalon’s services to develop models for business, science or economics.

Ben Vigoda, who sold Lyric Semiconductor to Analog Devices in 2011, described his approach as one based on comparing stories rather than numbers. Today’s deep-learning approaches will remain useful for tasks including images and video recognition, he said.

“Close to sensor, [traditional] deep learning is still best,” he added.

For now, Gamalon aims to keep its algorithms proprietary and sell cloud-based services. “Almost everyone has a problem with unstructured data—Gartner claims that 90% of all data is unstructured,” he said.

 
Next: Graphcore describes software behind AI processor »