So far deep learning models depend on powerful computing resources.
Do you think deep learning architectures have the capacity to be incorporated in the resource constrained devices or edge devices such that we can have some sort of on-device analytics?
If so, which method would be best suited for that? Any research on this area?
Deep learning gives you an answer/pattern. But that answer could also be found in through other ways. The research challenge in IoT is to replace deep learning with other simple light weight tools such that you can still get the results that deep learning gives. A strong understanding of statistics will cause less dependence on ready to use machine learning models which are kinda brute force at times.
Hi Mehdi Mohammadi,
We have just published a paper related to this very subject, but in respect to gaining practical access to deep learning resources through computational offloading to a server over a network. We present both a mathematical and laboratory-based experimental model to prove the concept and describe how it can be utilised for energy-constrained IoT platforms.
The paper is due to be published in Elsevier "Simulation Modelling Practice and Theory" and is titled as follows:
"Computation offloading of a vehicle’s continuous intrusion detection workload for energy efficiency and performance"
You can access the paper from my list of publications - I hope it is helpful to you.
Kind Regards,
Ryan
Sure you can. I just trained an AlexNet on a server then took the model and implememted it on a respberry pi 2. The classification on the pi take less than 100ms. We also trained a GoogleNet and tried it on the pi, it took 30 seconds per classification. Now we are upgrading to pi 3 with gpu capabilities.
Hi George,
Have you published any work on the above? I'd be interested to read it.
We used the Keras/TensorFlow package on a robotic vehicle, training remotely and running the classification both locally and remotely. Our work showed the detection (classification) latency was faster given the round-trip delay over the network as opposed to local classification on the vehicle; in this case it was not a PI, but ha access to an Intel Atom processor. As the deep-learning complexity increases, the classification time increases symptomatically of course and therefore local classification becomes some what impractical, especially if the IoT device performs actions in real-time such as a vehicle. That is, detection latency can result in cyber-physical problems. Depending on the application, however, local classification still seems very suitable if the resources are available.
If you get a chance please read our paper: "Computation offloading of a vehicle’s continuous intrusion detection workload for energy efficiency and performance" - I'd be interested to hear your feedback :)
Yes , We can have deep learning architectures in IoT enabled devices . We can have a offline training for the deep learning models and we can export the optimized parameters to the IoT devices . We have tried this on Pi2 for image classification and it works fine . Parameters reduction with maintaining the performance of the model is always challenging and we have to find a trade off for this .
There are actually a lot of research in this area, especially on quantization approaches and also using low-precision math. One place to get started is to take a look at the last ICML 2016 workshop regarding "On-device intelligence": https://sites.google.com/site/ondeviceintelligence/icml2016
Sure, deep learning architectures can be found in IoT enabled devices
- Badges
- Science topic