Googlers Create High-Power, Low-Footprint On-Device AI

Google Logo 2018 AM AH 2 new AH

In Short: Two AI researchers from Google recently put out a research paper detailing their creation of an advanced text classification AI that can run offline, even on devices with low specs like smartphones. The pair’s AI creation was able to achieve 86.7% accuracy on a simple data set, and 83.1% accuracy on a more complicated and multifaceted data set. The tests were all done locally, as well as on-the-fly training. Using self-governing neural networks helps to keep the footprint low and eliminate the need for the cloud, while the high accuracy and quick training are helped along by loss-defined ground truth, an algorithmical tool that modifies the model’s base, observed knowledge on the fly based on provided validation. It’s a form of supervised learning that takes minimal human involvement and can very quickly teach an AI to learn on its own. In this case, the AI was used to classify blurbs of text, sorting them based on whether they may be a sentence, a phrase, a  mathematical equation, an identification number, or a range of other things.

Background: Looking through the research paper, there are AI conventions first created all the way back in 2002 in use here, all centered around the core conceit of the aforementioned self-governing neural network. Essentially, you only have to guide such an AI until it figures out what you’re trying to make it do. From there, it learns everything else it needs on its own through trial and error. Given that this AI uses a loss-based ground truth, its baseline for calculation and sorting becomes more firm with each training run, making it less and less resource intensive as it continues to learn. As with most learning AI programs, how accurate this one becomes depends mostly on how long you give it to learn and how much material you expose it to while it’s in the learning process. In this particular case, an experimental AI that did not see real-world usage getting near 90% accuracy can be considered very impressive.

Impact: In a time where almost every flagship smartphone out there is shipping with onboard machine learning and neural networking features, an advancement like this could pave the way for just about any low-power friendly AI solution one could think of to make the jump to mobile devices. This is doubly true for devices that sport a separate coprocessor devoted entirely to AI functions, such as the one inside the Qualcomm Snapdragon 845 that powers flagship phones like the Samsung Galaxy S9 and Google Pixel 3. Text classification may not sound like much, but it has potential all its own, especially when it comes to things like live translation, UI navigation, and projects concerning helping the disabled navigate their devices. More exciting is the potential inherent in the combination of a self-governing neural network and loss-defined ground truth. This combination could easily lend itself to a wide range of other applications, from image classification to video game enemy AI, and its ability to run on lower-powered devices could mean that mainstream onboard AI finally becomes a reality. In places where expensive hardware or a constant internet connection could be hard to come by, this may be the development that finally starts bringing AI their way.