Design Methods for heterogeneous DNN
There are hundreds of use cases in the real world that require AI solutions through Machine Learning or Deep Neural Networks. With latest technical advancement, it takes practically very less time to build a vanilla DNN (Deep Neural Network ) engine, but the real challenge is about training them for the specific use case in hand. Except for those general purpose use cases for which massive amount of data are available in the internet, problems are for those specialized use cases for which data sets may be missing or they are not fit for purpose. So the challenge is to create training datasets for such use cases, and this task requires ingenuity and expert inputs for various behaviors and outcomes.
Once such inputs are made available, the next task will be to use mature tools or custom built data generation scripts to create massive datasets that mimic real world input data. Once such datasets are created and validated, the DNN can be trained. Such training methods are very common and require deeper understanding of the data being generated and the expertise to uniquely identify and generate data for every possible scenario that may occur.
Expert Judgement and Experience…
Expertise in such data set generation methods helps one to train networks to handle exception scenarios for which real data may not exist. But it is well known that training DNN requires massive amount of data for each scenario including exception cases, ability to generate large datasets for exception scenario will require expert judgement and insights. For example, training a DNN to identify minuscule fraud transactions within a large set of valid transactions require specialist skills to create different data sets of unique fraud transaction flavors. Such types of challenges do exist in production grade DNN deployments and it will be an iterative process that can take many months until a robust DNN based system can be built for such use cases.