Your browser is not supported. Please update it.

9 January 2014

TeraDeep’s image recognition to tempt the likes of Apple

Researchers at Purdue University are developing a neural networking chip which can power artificial intelligence applications such as image-recognition in mobile devices. The company TeraDeep has been formed around the idea and its hopes that Apple, Qualcomm, and other manufacturers will make use of it in upcoming handsets and devices.

Using deep learning, a series of algorithms that attempt to learn how to recognize images, the Purdue lab demonstrated how the chip can work with ‘face tracking, gesture recognition, augmented reality, and location-based services.’

Eugenio Culurciello, a professor involved in the project, hopes to use his company, TeraDeep, to commercialize the technique. By developing a coprocessor that uses neural-networking, the lab was able use a smartphone in their implementation. Google used 16,000 processors in their Stanford-led neural-network demo, which viewed and analyzed Youtube videos identifying common images, including cats.

A self-taught neural network does not need to be manually configured with labeled data. Adding such labels consumes time and energy, but the network can be left to learn independently, as the Google one did, which does not require further human input. Such independence led to Google’s network learning to recognize cats, through analyzing Youtube stills, to the delight of research fellow Jeff Dean.

He said ‘our hypothesis was that the neural network would learn to recognize common objects in the videos. Indeed, to our amusement, one of our artificial neurons learned to respond strongly to pictures of cats. Remember that this network had never been told what a cat was, nor was it given even a single image labeled as a cat.’

The results of Google’s experiment were encouraging, especially given the scale at which it was run. With only a billion connections, compared to the human brain’s 100 trillion, the network was able to identify common images amongst thousands of possible results. If applied to speech, text, or images, self-taught neural networks could prove very useful.

If the Google experiment had the equivalent power of a newborn brain, the potential of a fully developed 100 trillion connection network could be revolutionary. Compared to previous machine-learning image-recognition systems, the Google experiment achieved a 70% improvement, although its overall accuracy was only around 16%. The improvements that may be possible for such a system in simpler implementations demonstrates the potential for such systems in M2M communications. This is the area that we see huge potential in, for IoT M2M networks.

Given that the Google system taught itself cats in three days, a much simpler network could be quickly implemented in industrial or commercial settings. Traffic management, plant machinery, production lines and the like could all benefit from an independent neural network controlling the operation. Truly smart systems could be left to their own devices, learning the most efficient system without human input post-installation.

TeraDeep and their like demonstrate that neural networks can be used in much smaller implementations than the more famous neural experiments. The use of a coprocessor can allow a smartphone to use extremely powerful image-recognition capabilities in handsets, able to interpret events independently rather than as a pre-programmed response to an action. Augmented reality apps would see the most noticeable improvement through better image recognition, but the possibilities for translation or gesture control exist. Although the implementation of such technology in user apps looks likely, the greater impact will be seen in industry as the number of connections between devices continues to rise.