Email
Password
Remember meForgot password?
    Log in with Twitter

article imageGoogle's machine intelligence can now run offline on smartwatches

By James Walker     Feb 10, 2017 in Technology
Google has explained how it managed to get advanced artificial intelligence running offline on smartwatch processors. The system, formerly confined to high-performance datacentres, now powers Android Wear 2.0's suggested message replies.
The added reply options are a significant new Android Wear feature. Not only do they make it easier to respond to text messages but they also mark a step forward in the development of machine intelligence. The replies use advanced AI algorithms but require no internet connection.
In a blog post yesterday, Google's research team detailed how it brought AI from server rooms to your wrist. The project started with Allo and Duo, Google's new messaging apps that use machine learning. The team behind the apps was approached by the Android Wear group and asked if Allo's Smart Reply could run on a watch. The answer was a resounding no.
Smart Reply for smartwatches
Recognising the opportunity to solve a unique problem, the team working on Smart Reply went back to the drawing board. Its work so far was thrown out as the focus shifted to building a new, lightweight machine learning architecture that runs offline.
Just getting it working without an internet connection would have been a significant achievement. The team went a step further though, committing to bringing Smart Reply to smartwatches. As this week's announcement shows, it eventually succeeded.
Improved match grouping
The system is based on a new grouping mechanism that's faster and more efficient than existing models. Previous attempts to use on-device AI have revolved around simplistic systems based on rule mappings.
These models can detect positive emotion in phrases like "I love this movie" but do not scale to understanding the diverse vocabularies encountered in real text message exchanges. Neural networks are required for this but they're demanding algorithms that won't run on a wearable's tiny processor.
Projection model
Google's new solution aims for a middle ground in between the two. It first groups similar messages by comparing them using vectors. A locality sensitive hashing system is able to reduce the number of potential matches from millions of unique words to a specific sequence of bits, allowing messages to be compared with each other. The information is then fed through a "projection model" that can predict probable replies based on previous learning.
The complete system is highly complex and explained in more detail in Google's blog post. The outcome is an AI that can be trained using the cloud but then essentially downloaded to devices. The phone, tablet or wearable can then use an enhanced form of the inexpensive mapping procedures formerly used for on-device AI.
"Excited about how well it works"
In use, the system works well with much greater accuracy than previous models. Google said it performs better than anticipated, leading it to continue expanding the algorithms in the coming months. AI is now baked into new Android Wear smartwatches, without any strings to the cloud.
"When we embarked on our journey to build this technology from scratch, we weren't sure if the predictions would be useful or of sufficient quality," said Google. "We're quite surprised and excited about how well it works even on Android wearable devices with very limited computation and memory resources."
The development could be a pivotal point in the next evolution of AI. Over the next few years, machine intelligence is expected to move out of the datacentre and into the devices around you. They'll be able to help you while offline and reduce the time taken to respond to queries.
Google is starting small with suggested message responses. However, the company said it's already using the model to build "completely new applications," hinting that on-device AI will do much more in the future.
More about Google, Ai, Artificial intelligence, machine learning, neural networks