-
-A lot of libraries and frameworks allow to integrate model directly into edge devices but require extra steps and boilerplate and do not always support all the PyTorch and fastai layers
-Instead wherever possible deploy the model itself to a server and have your mobile or edge applications connect to it as a web service
What could go wrong now?
-There may be data that our model sees in production that is very different from what is saw during training (out-of-domain)
-Type of data our model sees changes over time and the type of risks it represents may change so much that the original training data is no longer relevant (domain-shift)
-Both are examples of a larger problem: that you can never fully understand all the possible behaviours of a neural network because they have far too many paramters
-Model may change the behaviour of the system it is a part of: in the presence of bias, feedback lookps can result in negative implications of that bias getting worse and worse (nigg)
Solutions:
1. First step is to use an entirely manual process with your deep learning model approach running in parallel but not being used directly to drive any actions with humans invovled in process should look at outputs and check whether they make sense
2. Second step is to try to limit the scope of the model geographically and time-constrained and have it carefully supervised by people
3. A helpful exercise prior to rolling out a significant machine learning system is to consider the question what would happend if it went really really well in other words, what if the predictive power was extremely high and its ability to influence behaviour was extremely significant? in which case who would be most impacted?