Difficulty level 2: Self-trained CNN on a custom dataset using TensorFlow and changes to start/end layers on the NN
Difficulty level 3: Custom rolled 12 layer CNN trained with with novel hand labeled data.
How many people to you think know the difference between these and the fact that each step is a magnitude or more harder than the last? That's not even getting into trying to apply research grade stuff.
Consider taking the side mirror from a Ferrari, putting it on a Ford Escort, and then pitching the Ford Escort as a race-car. The Ford Escort may be a great family car, but it is not a sports car. People who don't know the difference might come to think of it as a sports car, and might even question the wisdom of spending money on a Ferrari.
Most off the shelf ML tools are not that powerful - and they don't make a difference in how you differentiate as a business. If you have a ML team and know how to build models from scratch with novel data, then you can iterate on your model and expand, something you can't do if you don't own the model.
talk about dreamland autoaim
Effort where it maximizes return, leveraging existing assets where possible, to solve problems for real customers, is what's valuable.
Newly built tech is a rapidly depreciating asset; the depreciation comes not only from the progressing state of the art elsewhere, but also in maintenance costs and lack of network effects. Open sourcing to offset the depreciation is risky too.
The ML boom (specifically DL) is so interesting because of the applications. Computers recognising images better than humans for instance.
Surely outside of academia its a means to an end like anything else. If Team A wires together some high level Keras components with 30 lines of code, and does something really cool with it, is it less valuable?
What, basic linear regression is not cutting edge enough for you?
At a hackathon recently, someone pitched their app as an "ML-driven app" though the only ML in there was some 1-line language translation feature they were consuming off Watson REST services for a tangential feature on their app.
Meanwhile, my submission actually used a self-trained CNN on a custom dataset using TensorFlow and changes to start/end layers on the NN. The image classification features were the core of the app and it wasn't something that was just a wrapper over an Off-the-shelf API. We actually tried multiple networks and went thru the trouble of parameterizing everything.
At the end, I wonder how many judges actually understood the difference in effort/value to the two attempts at ML.