Deep learning in real time embedded systems
Deploying machine learning in real time embedded systems often comes with the constraint of reducing the latency, the memory footprint and the power consumption while achieving the same accuracy. BUT: in certain applications (e.g. Functional-Safety critical) you need strong guarantees that what your neural network infers is the right thing (i.e. a 20 km/h limitation is not a 50 km/h limitation!). You might think that in order to achieve good accuracy you need as many fractional bits as possible.