Tradeoff Quiz

Imagine you are tasked with choosing an initial DNN architecture for rapidly prototyping an upgrade to an existing mobile robot with the following constraints:

The current robot platform is a Jetson TX1 and has already been characterised and benchmarked with some well-known DNN architectures in the following graph.
The Jetson TX2 with an upgraded software stack has tentatively been selected for the upgrade with an expected 3x speedup in inference over the speeds in the benchmark graph.
The new project requires 70% accuracy and an inference speed of no less than 50 fps.
The lowest acceptable energy consumption is desired, because the longer the robot can run on battery, the better the customer experience.

Which architecture is the best candidate for the new platform and why?

GoogLeNet because it can run inference at 50Hz and it’s accuracy will be improved in the upgraded platform.

ResNet-34 because it will be faster than 50 fps in the new system and has an accuracy greater than 70%. It’s also has the lowest power usage of the acceptable candidates based on the assumption that inference time and operations count are linear.

VGG because it will run at 50 fps with the speedup and it is well known for its lean operations and parameter profile.

ResNet-50 because it will be faster than 50 fps in the new system and has an accuracy greater than 70%. It’s also has the lowest power usage of the acceptable candidates based on the assumption that inference time and operations count are linear.

None of the above

SOLUTION:

**ResNet-34** because it will be faster than 50 fps in the new system and has an accuracy greater than 70%. It’s also has the lowest power usage of the acceptable candidates based on the assumption that inference time and operations count are linear.