I am wary when people wax lyrical about all of the ways they love using machine learning on data. It makes me nervous because i worry that they have a hammer and can't wait to use it on anything vaguely nail shaped.
Machine learning is an area where you need to be able to produce results. Fake it ‘til you make it isn’t going to cut it for long. Either these people produce something that works, or they don’t.
Having to produce results is one thing. Mindlessly throwing tensorflow/pytorch at problems is an entirely different problem.
It's like those front-end devs who mindlessly insist that they need to use heavy javascript frameworks with convoluted build processes such as React/Angular to churn out a static web page with a couple of paragraphs and images.
You can certainly produce some plots and numbers, and possibly even plots and numbers that look good to your boss/clients/investors. The (multi)million dollar question is whether those numbers are actually meaningful. I think this is where a lot of ‘data science’, both in industry and academia, falls down.
Some state-of-the-art models don’t even generalize to test sets drawn from the same database, let alone similar data sources or the actual business problem. Unless you run a pet shop, telling breeds of dog apart, a la ImageNet, is probably not your goal.
Well, no. ML solves the classification problem, not the prediction problem.
E.g.: The "is this a cat picture" problem is effectively solved, but we _still_ can't reliably predict something as primitive as a simple binary proportion.