Despoisj

I that this addresses a threat model that we often neglect in AI safety talks.

It's not about a godlike self-improving AGI destroying the world (yet), but a more virus-like process that is surprisingly close to being possible today.

Not a malicious sentient agent plotting something evil, but just the AI version of a computer worm, replicating and hiding, hoping to move closer to its initial goal.

(Hopefully not to disassemble carbon-based lifeforms to make more paperclips.)

It's hard to see how we could effectively get rid of such a polymorphic virus, granted that the compute and weight size are small enough.

I made a 6 sided square flexagon puzzle for my dad, it's surprisingly simple yet can take forever to solve.

If interested, I've spend quite some time looking for more square flexagons, and made one with 12 and another with 14 faces (along with a puzzle version)! Tricky to build properly but stunning.

Thanks for the feedback!

As a former Kindle developer, I can say that most of what's mentioned in this article are metrics used to understand how the features are used (bookmarks, highlights, dictionnary, etc.), how much they are used, and in which country. This allows the teams to focus on features that are actively used, and sometimes lead to discontinuing features that see little to no use. Hope that helps.

During the interview for my current job (Machine Learning & Deep Learning Scientist), I didn't get any theoretical questions on machine learning, but only questions on my projects.

I didn't study machine learning per say, I followed dozens of online courses but no diploma in this exact field, and my last job was in traditional software engineering.

Despite this, I managed to do two small projects (one of music genre recognition [1] on my own iTunes library, and one of eye motion recognition [2]). For both I had published the code on my GitHub, and written a small Medium article to expose the problem I was trying to solve, the data available, the issues I faced and how I overcame them.

I personally believe that this is all the recruiter wanted to know, and why I encourage interns and candidates to work on their side projects.

References of the projects: [1] https://github.com/despoisj/DeepAudioClassification [1] https://medium.com/@juliendespois/finding-the-genre-of-a-son... [2] https://github.com/despoisj/DeepEyeControl [2] https://medium.com/hackernoon/talk-to-you-computer-with-you-...

What about using the "-" sign to filter results instead of relying on complex language understanding?

=> "shirt -stripes" works pretty well on google at least

This is a very naive question, but considering the artifacts you have on the GAN-generated images, you are trying to generate the whole "palette image" at a certain resolution.

For which reasons did you not use a 5x1 pixel RGB image as the target, that would have been orders of magnitude faster to train?

Thanks!

I'm not sure what you mean, but I have chosen to store slices on the disk so that I could still take a look at them, and not store the data only in numpy arrays. That could be optimize for a better processing time!

Thanks for the awesome ressources ! :D

The slice size is indeed quite arbitrary, and knowing the BPM would help, but isn't reliable either (various tempos, rubato for classical etc.)

I didn't intend to go that far in the human genre recognition parallel, but thanks for the references ! Good job on the script too

Thanks for your insights! I agree that log/mel spectrograms could be even more detailed and effective, and could be used with the SoX patch discussed here https://sourceforge.net/p/sox/feature-requests/176/.

Wow, that's an in-depth analysis of the task ! Thanks for sharing

Thanks for your inputs, it's true that we can use convolutions on raw waveform, however the main reason I've used a spectrogram was to work on precomputed relevant features as highd pointed out, instead of running the convolution on lots of data.

Yeah that's a problem I know - I used to make some Electro/Dubstep/Trap music - and I feel people will always disagree with the genre you pick anyway.

Thanks for the kind words, I'll take a look !

It's true that having very different genres helps the model a lot. It would be much more difficult to distinguish between closer genres, especially when people don't really know which is which and argue all the time about it.

Haha, I should have titled the article "How to build an internet music-genre-troll-bot"

It's true that the validation set only contains genres I used for the training. I'll try this out this evening ;)

Haha that would be awesome ! I guess we'd need a lot of data, and probably use much more detailed spectrogram (time-wise and frequency-wise).

Sorry if the 5 slices are misleading it was only for readability, the average song has 70 slices, which are all classified and used or voting.

Yep you got it right, except the voting system adds tons of reliability because we cannot trust the slice classification (2.5s) too much.