3
$\begingroup$

I have encountered this pattern for a long time (5+ years). So many professionals come with an interesting domain-specific problem, and they demand using state-of-the-art deep learning models: take it or leave it.

I understand that technology advances faster than ever, but I am still missing the point. Indeed, I often propose using simpler, traditional ML models over complex ones because, for example, in an MLOps scenario is better to start with simpler models and then move to more complex ones. But, unfortunately, business experts often seem to be disappointed with such kind of proposals. Moreover, there are many reasons to prefer classic ML, which I often use to motivate the latter.

$\endgroup$
1
  • 1
    $\begingroup$ It's commonsense if you have a ML or stats background. From the outside, recent DL results seriously look like magical fairy dust. It's either that, or blockchains. $\endgroup$
    – maxy
    Jan 28, 2023 at 7:06

2 Answers 2

2
$\begingroup$

Short answer: because they are not experts in ML (and they should not be, otherwise they won't be asking), but are bombarded by buzzwords e.g. AI, blockchain, ChatGPT.

Do you have any friends who put their whole lifesaving in cryptocurrency without any idea what they bought? Same thing. This is human nature.

$\endgroup$
2
  • $\begingroup$ Yes, that is true indeed. Nevertheless, I feel bothered when non-ML experts say those fancy words just because they heard them in the news. I recall one day that a person thought cloud computing was performed in the cloud; it seemed like a joke, but it was not. I will keep standing my ground, fighting against such useless arguments and presenting simple, yet effective, solutions. Thank you. $\endgroup$
    – Eduard
    Feb 3, 2023 at 18:20
  • $\begingroup$ remind me of this $\endgroup$
    – lpounng
    Feb 4, 2023 at 5:28
2
$\begingroup$

There is another factor not yet mentioned.

Classic ML techniques potential is usually capped, we know already the limitations and the increased sophistication required to improve accuracy. Of course this is just our (perhaps faulty) perception of them, due to the slow improvements of the field.

DL techniques maximum potential is still unknown and as such is it somewhat rational to bet on them. Indeed the bitter lesson is that being able to scale well with the improvements in compute tend to be superior to hand engineering over the long term.

The above reasoning similarly explains why new technologies are "exciting".

$\endgroup$
1
  • 2
    $\begingroup$ That said, I'm not advocating to go DL first! As the OP said, the best practice is to start simple and increase complexity only if needed. $\endgroup$
    – Rexcirus
    Feb 22, 2023 at 13:39

Not the answer you're looking for? Browse other questions tagged .