AutoML's role in AI


AutoML is becoming something that even non ML practitioners can use, develop and deploy ML products into production. What would be the role of ML engineers when organisations lean towards an AutoML?

There are gaps in AutoML which i thought still exists such that we can’t rely 100% on it, though the industry is certainly closing its gaps.

  1. Explainability. Having experience with some AutoML tools, my general feel is that it tends to go for the best outcome without much consideration on complexity. This may put a strain of explainability and interpretability on models outputs.
  2. Tasks. Areas such as Reinforcement Learning and Agent based learning are still not a common thing in AutoML products. Less general tasks such as Person ReID and object tracking don’t typically come packaged in AutoML products. So ML Engineers are certainly still required here.
  3. NAS. As fantastic as it sounds, the AutoML solution still need to keep up with the latest in development, and this can take time. ML Engineers can however jump straight into it.
  4. Data Centric AI. AutoML works on whatever you feed it. I haven’t seen one that says, ok you need to improve that data quality. This is something which humans can still pick up generally well.

For discussion.

Thank you

1 Like

Like the general programming world. Idk to give my cents bc of my very early background in .NET world.

So I did .NET task in Visual Studio .NET. You know what that there were a lot of components and controls that you could do via drag-drop to connect dataset or report tables. However, for details I still had to work with coding or classical design pattern (GOF ie.)

What Idk to say is that a lot of parts in the area of ML or DL need knowledge, especially in tuning the models or improving training performance.

I don’t know what and how good the cloudy platforms can provide, but components like learning-rate (⍺), ε, β, or other learning parameters like in BatchNorm, bla bla bla , they are not just a work of in UI :stuck_out_tongue_winking_eye:.

To shorten, the AutoML is a tool. The AutoML gives me a quick and half-done model, like using Norm Equation for a small dataset. I need them to find out a prototype and even give me a visualization approach.

I hope that it helps.