Deep Learning Possibilities For Mobile
Source: Tina Rose
MIT recently presented their new 168-core chip Eyeriss that has deep learning capabilities for mobile. The excitement at the International Solid State Circuit Conference surrounded the possibilities for this new chip in mobile applications as well as the Internet of Things. Eyeriss offers the possibility of holding features like facial, object, and sound recognition in the palm of your hand.
A New Chip from MIT
Eyeriss, a new 168-core chip, was presented by MIT recently at the International Solid State Circuits Conference in San Francisco. The team of researchers impressed everyone with its deep learning capabilities and ability to run powerful AI algorithms locally. It will be 10 times more powerful as a mobile GPU. Energy friendly, the processor runs AI algorithms locally rather than uploading to the Internet.
In the past, different services like facial, object, and sound recognition relied on the cloud, however Eyeriss will have the capabilities to do this on its own. Because each of its cores has its own memory bank, there is a potential to take mobile AI to a whole new level of products. Decisions can be made at the cellular level – the need for a Wi-Fi or cloud connection will be minimized.
The Power of Eyeriss
Eyeriss is being developed so devices can make decisions and do more things without human intervention. Eyeriss is a self-contained deep learning system that will be able to tap its own memory to recognize sounds, faces, and objects. Its uses are endless in phones, self-driving cars, robots, drones, and wearables.
Complex neural networks are typically required along with vast computing resources and servers. Centralized memory systems of GPUs and CPUs power the complex neural networks required by deep learning systems.
future
MIT says its chips would require a fraction of the resources, and is 10 times more power efficient than a mobile graphics processor.
It works because Eyeriss minimizes the amount in which the cores need to exchange data with distant memory banks because each core contains its own memory bank. This chip reduces repetition by breaking down tasks among all of its 168 cores. It can be reconfigured to different types of neural networks and compression preserves bandwidths.
Deep Learning Without the Cloud
The idea is that by minimizing the need to exchange data with distant memory banks the Eyeriss chip and its devices will rely less on the cloud and will be able to make decisions locally and act accordingly. By compressing data before sending it to individual cores – each core can communicate directly with its neighbor – so by sharing data, they don’t have to continually route through main memory.
Eyeriss has the capabilities of bringing self-contained AI capabilities to most of our devices with the processes happening locally on the device itself. Since Wi-Fi or cellular connections won’t be needed, the cloud is not needed for image recognition or decision-making. The key to the chips abilities is a circuit that allocates tasks across the 168 cores. In its local memory, the core stores both the data manipulated by the nodes its simulating but also the date described by the nodes themselves. The memory can be configured for different types of networks, and then applied at the start of running the application.
Deep learning capabilities on our mobile devices brings a lot of possibilities. In areas where Wi-Fi is not available you could still access services, use recognition software and make decisions all without accessing the cloud. For privacy reasons alone, you may want to process your applications locally. Another advanced technology we could look forward to is that with MIT’s chip, self-driving cars can operate in remote areas where cellular connections are not available.
“Deep learning will make you acceptable to the learned; but it is only an obliging and easy behavior, and entertaining conversation, that will make you agreeable to all companies.” James Burgh
| }
|