It feels like the answer to this at my last job was: “All the time.”
We worked a lot with time-series data. I would always make a statistical model as a “baseline,” train some version of an RNN, and immediately realize my baseline would have a better cost/metric ratio.
At the end of the day, at least in my experience, it always came down to cost. If a simpler non-DL model could do 90% of the job at 50% cost - it’s hard to convince business teams why we should shell out the dollars for the better model.
I am totally with you on this one @chris
I have done a lot of work on stocks data and algorithmic trading and i have always seen non-DL models working better than DL models with less compute resources.
To be honest in computer vision DL is usually the way to go - using old school feature extractors have always proven to be less effective than a good DL embedding.
That being said I CAN speak to times where people probably didn’t need a model at all because doing it manually would be faster hahaha. (The model couldn’t take the human fully out of the loop - and was slower than the human would’ve been operating manually).
Last time…hmmm…my strategy for solving problems is to use an easy solution first, either to produce a proof of concept quickly or eliminate simple techniques that don’t provide results.
All to say, me using my deep learning powers is a last resort, or should I say second to last (StackOverFlow is the last resort haha).
All jokes aside the last time I had to use a dl technique to solve a problem when something easier could be used would be at my previous job where I had to do some simple activity recognition.
The lead wanted to use dl for something that could be solved using a trivial algorithm and heuristics.
I spend months collecting data to train on specific moves, which was absolutely a waste of time.
I feel like there’s a good lesson implicit there. How do you respond in a situation when a lead insists on using an overly complex technique? Maybe an indication they shouldn’t be in that role in the first place?
Curious about the non-DL techniques that you used for activity recognition. Would like to hear more about that.
You are right, individuals that don’t understand applied machine learning should not be in a position to lead ML teams, sounds very obvious, but you will be surprised at the fact that a lot of people believe that individuals that are managers or lead within a technical division that’s non-ml related, can also have a successful time leading an ML team.
That assumption is very wrong and dangerous to an organisation.
A good Product manager or DevOps manager doesn’t make for a good ML team lead
In regards to the non-DL technique to activity recognition.
For proof of concept application, it is usually quicker to find the key joints associated with an action, get the coordinates and record certain distance measures and snapshot poses to match against incoming poses in real-time.
Much faster, saves time and then thereafter you can adopt more complex DL techniques, like using an LSTM or something