[okabat]

The fastest way to make progress on these business questions is often to build hacky MVPs that look like they're doing something smart, but behind the scenes are powered by humans or dead-simple algorithms, and get them in front of customers ASAP.

I recently joined a seed-stage startup solving a business problem via audio analysis in the manner I described above. I'm not spending much time doing ML yet, but I'm banking on my belief that we're solving a valuable problem (customers want to buy our hacky MVP) and that ML can and will be needed to scale our solution. By deeply understanding the customer as a first step, I think the ML systems we build will be business critical and enduring. Time will tell

[cirgue]

> dead-simple algorithms

The most successful models I've ever built have been logistic regression models. If you can rephrase your problem in a way that's amenable to run-of-the-mill statistical techniques, you can frequently achieve much better results than you can with 'deep learning'.

[madcaptenor]

If you're going to do logistic regression, at least call it a single-layer neural network with a sigmoid activation function.

[inputcoffee]

So true. I have had to do this.

I had a pretty good regression model but it was not taken seriously. So I wrote it using "a neural network in TensorFlow" and the next thing you know the whole company is asking me how it works and what it does.

[cirgue]

This feels so dirty, but these kinds of tricks work. Your stakeholders get to participate in the titillating fiction that they are on the bleeding edge of technology, and you get to deploy a scalable, explainable, and (hopefully) high-performing solution. That's 95% of a win.

[zebraflask]

I'm totally "borrowing" that for later use.

[bllguo]

Amen.. I like to use exotic techniques just out of intellectual curiosity and to put my education to use, but ultimately a basic linear regression is all people want! Ease of interpretation is paramount.