Jan Hosang
@janhosang.bsky.social
Tinkerer, dog-lover, MLbro, Computer Vision dude.
I know Google doesn't have the reputation, but I think the commitment of blurring SV and aerial data on request predates GDPR by years.
And unfortunately those requirements bleed into the model weights, too.
And unfortunately those requirements bleed into the model weights, too.
November 5, 2025 at 7:15 AM
I know Google doesn't have the reputation, but I think the commitment of blurring SV and aerial data on request predates GDPR by years.
And unfortunately those requirements bleed into the model weights, too.
And unfortunately those requirements bleed into the model weights, too.
Oh boy do I wish that was the reason. SV data is super sensitive and we have to comply with takedown requests.
November 3, 2025 at 6:42 PM
Oh boy do I wish that was the reason. SV data is super sensitive and we have to comply with takedown requests.
I believe it can also mean that it explains the difference between Y% of the pairs of individuals (instead of Y% of the difference for each pair). More generally, I think it can also mean it explains A% of the difference for B% of the pairs, s.th. A*B=Y.
August 18, 2025 at 2:11 PM
I believe it can also mean that it explains the difference between Y% of the pairs of individuals (instead of Y% of the difference for each pair). More generally, I think it can also mean it explains A% of the difference for B% of the pairs, s.th. A*B=Y.
It's still magic even if you know how it's done.
- Terry Pratchet
- Terry Pratchet
July 3, 2025 at 5:20 PM
It's still magic even if you know how it's done.
- Terry Pratchet
- Terry Pratchet
The hidden layer is intriguing, but I cannot imagine really enjoying the normal room building mechanics, so I thought I'd ask before taking the plunge. Thanks for answering!
April 21, 2025 at 12:16 PM
The hidden layer is intriguing, but I cannot imagine really enjoying the normal room building mechanics, so I thought I'd ask before taking the plunge. Thanks for answering!
Wow! Do you know why? What a marketing scam.
March 22, 2025 at 7:04 PM
Wow! Do you know why? What a marketing scam.
I agree. That the highest profile institution gives half pay for >100% work is not ok. And not taking allegations seriously is unacceptable, particularly as a publicly funded institution.
March 16, 2025 at 10:37 AM
I agree. That the highest profile institution gives half pay for >100% work is not ok. And not taking allegations seriously is unacceptable, particularly as a publicly funded institution.
Arrested Development Dead Dove GIF
ALT: Arrested Development Dead Dove GIF
media.tenor.com
February 21, 2025 at 3:54 PM
How did you know what to change to encourage emergence?
February 21, 2025 at 1:35 PM
How did you know what to change to encourage emergence?
That probably depends on the task, right? Have you done a specific experiment? For needle-in-haystack I can imagine that. For correspondence tasks not really.
February 14, 2025 at 3:21 PM
That probably depends on the task, right? Have you done a specific experiment? For needle-in-haystack I can imagine that. For correspondence tasks not really.
If you believe you can cram all information into one vector, then maybe it's not such a stretch to believe you can go over the points in random order? Though I have a hard time buying that.
February 14, 2025 at 2:25 PM
If you believe you can cram all information into one vector, then maybe it's not such a stretch to believe you can go over the points in random order? Though I have a hard time buying that.
I mean bsky feed of course.
February 14, 2025 at 2:05 PM
I mean bsky feed of course.
Maybe we should make a bsky list for paper discussions? My feed is a little messy.
February 14, 2025 at 2:04 PM
Maybe we should make a bsky list for paper discussions? My feed is a little messy.
The surprise of a paper's results should be given more weight also while reviewing. We overindex so hard on novelty or flat numbers, instead of insights.
February 14, 2025 at 2:02 PM
The surprise of a paper's results should be given more weight also while reviewing. We overindex so hard on novelty or flat numbers, instead of insights.
If we use permutation invariant models for sequence modeling, what would be more fitting than using sequence models on unordered sets? /s
February 14, 2025 at 1:58 PM
If we use permutation invariant models for sequence modeling, what would be more fitting than using sequence models on unordered sets? /s
I've been fascinated by the sound Hammond organs (not sure if that's the right term) make when playing more tones at the same time. Maybe the overtones interact in interesting ways? But I love this growl, too. It's great when key instruments get the expressive range of electric guitars for example!
February 13, 2025 at 6:55 AM
I've been fascinated by the sound Hammond organs (not sure if that's the right term) make when playing more tones at the same time. Maybe the overtones interact in interesting ways? But I love this growl, too. It's great when key instruments get the expressive range of electric guitars for example!
How does the low E get so dirty? Is it just a timing based effect?
February 12, 2025 at 2:39 PM
How does the low E get so dirty? Is it just a timing based effect?
"share the same last braincell". I'm stealing this one.
November 28, 2024 at 7:05 PM
"share the same last braincell". I'm stealing this one.
Nice to dog-meet! Morning walks need to be foggy.
November 18, 2024 at 8:11 PM
Nice to dog-meet! Morning walks need to be foggy.
Thanks, Chris! That's one I'm definitely aware of. I'd be more interested in semantic features than coming up with the points only.
November 17, 2024 at 3:47 PM
Thanks, Chris! That's one I'm definitely aware of. I'd be more interested in semantic features than coming up with the points only.