Towards Wearable Cognitive Assistance

Motivation:

Wearable devices can help people by providing cognitive assistance by recognizing images or sound and processing them. But the wearable devices have limited computational power and energy to process the data acquired. So the solution for this problem could be computation offloading to the servers and get the data processed in much faster rate.

Main points

• Wearable devices can act as multiple sensor devices to recognize objects, sounds and location as per need. Eg: Google glass can perform all the mentioned tasks.
• After capturing the image or sound, it has to be processed. That would require complex computer vision algorithms or machine learning algorithms to be run on the data. Wearable device lacks such power and space to perform such computation.
• So the data is transferred to the server for processing. If a cloud server like Amazon EC2 is used, it will incur more latency which is not acceptable for personal assistance kind of applications.
• To make RRT shorter, cloudlets are used which are available in nearby location that will perform the computation on behalf of the wearable device.
• Server architecture is multi-tiered which consists of multiple VMs running in parallel which will take up some task from the wearable device to perform the computation parallely to complete faster.
• There is a control VM which receives the data from wearable device. It understands the context and assigns the tasks to the cognitive VM Engines to perform the required task like face recognition,etc. Control VM also keep track of the load and job status on the available cognitive VM Engines for coordination.
• When cloudlet is not available in nearby location, computation can be moved to cloud server in internet is available or to your laptop or mobile phone which you are carrying and communicate using Bluetooth.

Comments:

Positive Points:

1. This paper leverages on the rapidly developing mobile to cloud offloading technologies for wearables which is a good idea.
2. It covers almost all possible aspects in offloading the computation from mobile to cloud, cloudlet, then to local computer or mobile to give seamless user interaction.

Trade Offs

1. Can the cloudlets or local computer contain global database of images or sound files. Most of the computer vision or ML algorithms requires large volume of image/sound data to classify or match. If not, it might give false positives as results. Is it always a good idea to move the computation to cloud?
2. This requires a highly sophisticated setup of cloudlets with pre installed VMs(Cognitive Engines) which may differ from application to application. Can the cognitive VM engines shared across applications and users?