Hacker News Reader

Comment by pool1892 | Hacker News Reader

how is this a new observation in any way?

I do not think it is supposed to be new. Everyone knows the flaws but downplay it by a lot. (negative proofs doesn't get funding and all that...)

I see this as a very funny way to recall what we always knew. Like a court jester reminding the king he was always naked.

This analysis is quite the strawman. It's a bit like taking a demo SQL table and basic select queries off an intro to SQL tutorial and claiming SQL sucks because the queries are inefficient and the tables can only store books and author names but not any other kind of data because that's what the example was. There is a lot more to SQL than what gets introduced in the intro to SQL tutorial and there is a lot more to computer vision nets than grabbing the network trained on a benchmark dataset and shoving completely different data at it.

>there is a lot more to computer vision nets than grabbing the network trained on a benchmark dataset and shoving completely different data at it.

Sure, you are right, but lots of people are wandering about saying different!

Interesting, maybe I'm just in my own filter bubble, but I never see that suggestion. After sampling roughly 15 articles found searching for things like "how to use inception net", etc I didnt find a single article suggesting you take a pre trained model based on image net and use it without modification on a completely different dataset. I am genuinely curious, do you have any examples of this?

Also, it may seem minor, but as someone who has been working professionally with neural networks since 2013 or so, there is a huge difference between saying "what inception net doesn't see" and "what inception net trained on image net doesn't see". I may grant that the author is simplifying for an inexperienced audience, but if that were the case, they should at least point out how such problems can be avoided, and have been avoided for at least 5 or 6 years now, if not longer.

Maybe the GP has a job that means that they can't comfortably give some specifics. However, literally every GCP and AWS customer engineer that I have been on a call with says "you can use these models out of the box for any application and they will just work". Also, all of my competitors are saying "ML is democratized you just plug the models in, the cost of your project will be 3 engineers for eight weeks from India/Ukraine/Brazil". The problem is that this is leading to unnecessary failures of the technology both because it means that many projects that should not be attempted are being attempted and are doomed to fail, and on the other hand it means that projects that could work if done right are failing unnecessarily. This is burning investment capital and good will all over the place. In my opinion educating the market is essential : any dolt can wire together a demo that does 80% of what's needed, sales people can use that to convince the customer that the outcome will be a project that costs in at 30% of the value and delivers 110% (they are sales, so maths is not important to them) but delivering that project requires technical insight, competence, structure and professionalism. All of which are in short supply.

(Original blogger here.) This isn't a research paper describing a novel contribution! Just a simple blog post with some observations I made with a Python library (Gradio) that I helped develop :)

Reply on news.ycombinator.com