Ask HN: Can ML help filter YouTube better | Hacker News Reader

Ask HN: Can ML help filter YouTube better

2 pointscrate_barre9 comments

I really need the following things filtered:

1. Any video using sentimental music (usually piano) to explain a story (helping someone etc, rescuing and a dog and shit).

2. Need someone to model those video with cadences where someone talks like a YouTuber, you know when their voice cuts right into the next sentence, sort of like removing frames from an action movie. They are getting to sound like that TV reporter archetype.

3. Identify the ‘please please like and subscribe’

4. Can we model the live ad reads and catch those?

I honestly don’t know anything about ML but I’ll literally sit through anyone’s Coursera (Will pay $$$) if you just walk through the concepts and how things like this can be tackled.

In other words, I’m suggesting the future of spam and ad blocking has to occur on this level, because there’s no other way.

View on HN

I agree with jstx1's sentiment. It would be easier to run a 3rd party service that users can submit information to a la SponsorBlock:

> SponsorBlock is an open-source crowdsourced browser extension and open API for skipping sponsor segments in YouTube videos. Users submit when a sponsor happens from the extension, and the extension automatically skips sponsors it knows about using a privacy preserving query system. It also supports skipping other categories, such as intros, outros and reminders to subscribe, and skipping to the point with highlight.

I’m trying to go beyond the ads. I want the automotans that are pumping out template videos with a certain script, voice attenuations, to be filtered out by an an actual automotan, as in, if you are going to pump these things out like a machine - then let’s have the machine just identify you and cancel you out.

I don't think it's practical and your examples are too specific. For the promotions (or auto promotions) there is SponsorBlock, it works pretty well for most well known channel.

For the rest, it's mostly creator specific. If a creator is abusing sentimental music in many videos, or I don't like the way he/she talks. I just hit "unsubscribe" and move on.

Not that it's important, but your post reminded me of a movie quote that I finally figured out is from Sherlock Holmes, Game of Shadows.

Moriarty:

> So you're not fighting me...so much as you are the human condition

There's two different questions:

- can ML algortihms do these things? yes, to all of your points

- can you realistically run such an ML service? only if you're youtube.

How practical is it to run some of this client side on the browser via WASM, with a few things offloaded to a server hit? Our computers are getting better. Think plug-in, think ad-blocker.

The difficult part is training a model like this - it's very expensive to train a huge model on video, you need lots of work and infrastructure, and you can't legally do any of this in the first place.

Why would it need that much training data? It takes me 3 or 4 videos to pick up the pattern. It’s super patternistic, and I almost have a hard time believing it would require that much training data. It’s almost like the Shazam app, I think you can literally compare the graph of the song and identify it quickly.

> Why would it need that much training data? It takes me 3 or 4 videos to pick up the pattern.

You're human being capable of understanding what you're seeing, ML is just computational statistics. You need to throw a ton of data at it to get accurate results.

You might be able to piggyback on pre-trained models via transfer learning in order to reduce the amount of data and resources needed to train a model for this task, though.

I don't think Shazam uses ML for song recognition, but instead hashing and fingerprinting via Fourier transforms.

Reply on news.ycombinator.com