Jan van Gemert
@jvgemert.bsky.social
Head of the Computer Vision lab; TU Delft.
- Fundamental empirical Deep Learning research
- Visual inductive priors for data efficiency
Web: https://jvgemert.github.io/
- Fundamental empirical Deep Learning research
- Visual inductive priors for data efficiency
Web: https://jvgemert.github.io/
Haha, well, I was also thinking about CoordConv, the paper that added position info to convnets :)
(Which, interestingly, seems to go against that finding that convnets can already encode absolute position 🤔)
(Which, interestingly, seems to go against that finding that convnets can already encode absolute position 🤔)
January 12, 2026 at 8:15 PM
Haha, well, I was also thinking about CoordConv, the paper that added position info to convnets :)
(Which, interestingly, seems to go against that finding that convnets can already encode absolute position 🤔)
(Which, interestingly, seems to go against that finding that convnets can already encode absolute position 🤔)
I've been also thinking about position embeddings :). I might investigate with a MSc student, although that is still a while before they start.. 😅
January 12, 2026 at 8:03 AM
I've been also thinking about position embeddings :). I might investigate with a MSc student, although that is still a while before they start.. 😅
Ahaa! "negative impact" is a continuous term, not discrete.
Now I understand why they were even threatening to desk reject one of my papers that was already desk rejected ;)
(We didn't make it in time so it was already out..)
Now I understand why they were even threatening to desk reject one of my papers that was already desk rejected ;)
(We didn't make it in time so it was already out..)
January 9, 2026 at 12:36 PM
Ahaa! "negative impact" is a continuous term, not discrete.
Now I understand why they were even threatening to desk reject one of my papers that was already desk rejected ;)
(We didn't make it in time so it was already out..)
Now I understand why they were even threatening to desk reject one of my papers that was already desk rejected ;)
(We didn't make it in time so it was already out..)
Oh, i found the "urgent! Risk of desk reject when co-author reviews not done" quite exciting ;)
January 9, 2026 at 11:37 AM
Oh, i found the "urgent! Risk of desk reject when co-author reviews not done" quite exciting ;)
Interesting! Even in a time where the last bastion of "knowledge based visual feature engineering" (ie: geometry) is replaced by learning (eg: VGGT) ?
January 7, 2026 at 8:27 PM
Interesting! Even in a time where the last bastion of "knowledge based visual feature engineering" (ie: geometry) is replaced by learning (eg: VGGT) ?
Does it have to be 3 years? 🤔
For a failure, I can recommend my own TPAMI 2010 paper on "visual word ambiguity", exemplified by:
- not end-to-end (feature engineering)
- grayscale only
- convexity assumption
- small datasets
- ... (?)
For a failure, I can recommend my own TPAMI 2010 paper on "visual word ambiguity", exemplified by:
- not end-to-end (feature engineering)
- grayscale only
- convexity assumption
- small datasets
- ... (?)
January 7, 2026 at 6:07 PM
Does it have to be 3 years? 🤔
For a failure, I can recommend my own TPAMI 2010 paper on "visual word ambiguity", exemplified by:
- not end-to-end (feature engineering)
- grayscale only
- convexity assumption
- small datasets
- ... (?)
For a failure, I can recommend my own TPAMI 2010 paper on "visual word ambiguity", exemplified by:
- not end-to-end (feature engineering)
- grayscale only
- convexity assumption
- small datasets
- ... (?)
I misread it as "chess cake" 😂.
Or should it be "check, cake"? (With the bishop 🙂).
You taught the 17y old to play well, nice position 🙂.
Or should it be "check, cake"? (With the bishop 🙂).
You taught the 17y old to play well, nice position 🙂.
January 1, 2026 at 3:44 PM
I misread it as "chess cake" 😂.
Or should it be "check, cake"? (With the bishop 🙂).
You taught the 17y old to play well, nice position 🙂.
Or should it be "check, cake"? (With the bishop 🙂).
You taught the 17y old to play well, nice position 🙂.
I'm catching up during the hiking holidays 🙂
December 23, 2025 at 8:09 PM
I'm catching up during the hiking holidays 🙂
Did we truly understand this ourselves? 😉
December 21, 2025 at 9:41 PM
Did we truly understand this ourselves? 😉
@davidpicard.bsky.social is doing it in Paris, with cvpr pronounced in a French accent 🙂.
December 13, 2025 at 5:47 PM
@davidpicard.bsky.social is doing it in Paris, with cvpr pronounced in a French accent 🙂.
Dutch nationalistic reading? 😉🇳🇱
December 8, 2025 at 8:04 PM
Dutch nationalistic reading? 😉🇳🇱
Wonderful! I like nearly all of his work 🙂.
As i also do for most of the work of Neal Stephenson :)
As i also do for most of the work of Neal Stephenson :)
December 7, 2025 at 2:16 PM
Wonderful! I like nearly all of his work 🙂.
As i also do for most of the work of Neal Stephenson :)
As i also do for most of the work of Neal Stephenson :)
Very interesting! I will have a look at those sources. Thank you.
I'm in machine learning (but not "AI" 😂) myself, and shortcut learning is one of the unsolved (practical?) problems in our field.
I'm in machine learning (but not "AI" 😂) myself, and shortcut learning is one of the unsolved (practical?) problems in our field.
December 6, 2025 at 7:37 PM
Very interesting! I will have a look at those sources. Thank you.
I'm in machine learning (but not "AI" 😂) myself, and shortcut learning is one of the unsolved (practical?) problems in our field.
I'm in machine learning (but not "AI" 😂) myself, and shortcut learning is one of the unsolved (practical?) problems in our field.
Definitely a language issue, my apologies.
I guess his point is: there's conflict between gains for science and gains for the scientist.
And that aligning them is an unsolved problem; in machine learning there's Bostrom's paperclip example 🙂. Also Goodhart's law.
I guess his point is: there's conflict between gains for science and gains for the scientist.
And that aligning them is an unsolved problem; in machine learning there's Bostrom's paperclip example 🙂. Also Goodhart's law.
December 6, 2025 at 6:57 PM
Definitely a language issue, my apologies.
I guess his point is: there's conflict between gains for science and gains for the scientist.
And that aligning them is an unsolved problem; in machine learning there's Bostrom's paperclip example 🙂. Also Goodhart's law.
I guess his point is: there's conflict between gains for science and gains for the scientist.
And that aligning them is an unsolved problem; in machine learning there's Bostrom's paperclip example 🙂. Also Goodhart's law.
Incentives are intended as rewards; but that having principals hurts personally.
Ie: to have principles and not "just go along with shady business" might hurt your career.
Eg: criticizing misconduct of famous/powerful people might get your next grant proposal rejected.
Ie: to have principles and not "just go along with shady business" might hurt your career.
Eg: criticizing misconduct of famous/powerful people might get your next grant proposal rejected.
December 6, 2025 at 3:48 PM
Incentives are intended as rewards; but that having principals hurts personally.
Ie: to have principles and not "just go along with shady business" might hurt your career.
Eg: criticizing misconduct of famous/powerful people might get your next grant proposal rejected.
Ie: to have principles and not "just go along with shady business" might hurt your career.
Eg: criticizing misconduct of famous/powerful people might get your next grant proposal rejected.
Very well deserved! Proficiat! 🙂.
Keep up the fight; I really admire your work and courage!
Keep up the fight; I really admire your work and courage!
December 6, 2025 at 3:06 PM
Very well deserved! Proficiat! 🙂.
Keep up the fight; I really admire your work and courage!
Keep up the fight; I really admire your work and courage!
One is a Quick Response code!
It has to do with image recognition, which symbolizes the visual deep learning roots of NeurIPS. It's black and white pixels; because before 2012 only few people used color.
I assume this will soon change to small round plastic thingies.. or some other tokens..
It has to do with image recognition, which symbolizes the visual deep learning roots of NeurIPS. It's black and white pixels; because before 2012 only few people used color.
I assume this will soon change to small round plastic thingies.. or some other tokens..
December 2, 2025 at 11:31 PM
One is a Quick Response code!
It has to do with image recognition, which symbolizes the visual deep learning roots of NeurIPS. It's black and white pixels; because before 2012 only few people used color.
I assume this will soon change to small round plastic thingies.. or some other tokens..
It has to do with image recognition, which symbolizes the visual deep learning roots of NeurIPS. It's black and white pixels; because before 2012 only few people used color.
I assume this will soon change to small round plastic thingies.. or some other tokens..
Ik ben zelf onderzoeker, reviewer, chair 🙂.
De journals eisen van ons de rechten, waarom hebben ze dan niet de plichten?
Ze kunnen kwaliteitscontrole wel (gratis) uitbesteden, maar waarom zijn ze niet eindverantwoordelijk?
De journals eisen van ons de rechten, waarom hebben ze dan niet de plichten?
Ze kunnen kwaliteitscontrole wel (gratis) uitbesteden, maar waarom zijn ze niet eindverantwoordelijk?
December 1, 2025 at 8:18 AM
Ik ben zelf onderzoeker, reviewer, chair 🙂.
De journals eisen van ons de rechten, waarom hebben ze dan niet de plichten?
Ze kunnen kwaliteitscontrole wel (gratis) uitbesteden, maar waarom zijn ze niet eindverantwoordelijk?
De journals eisen van ons de rechten, waarom hebben ze dan niet de plichten?
Ze kunnen kwaliteitscontrole wel (gratis) uitbesteden, maar waarom zijn ze niet eindverantwoordelijk?
I also enjoyed "Jonathan Strange & Mr Norrell" from her :)
November 30, 2025 at 6:20 PM
I also enjoyed "Jonathan Strange & Mr Norrell" from her :)
Waarom zijn journals niet verantwoordelijkheid voor de kwaliteit van wat ze verkopen?
Ze krijgen tenslotte wel alle inkomsten.
Ze krijgen tenslotte wel alle inkomsten.
November 30, 2025 at 6:14 PM
Waarom zijn journals niet verantwoordelijkheid voor de kwaliteit van wat ze verkopen?
Ze krijgen tenslotte wel alle inkomsten.
Ze krijgen tenslotte wel alle inkomsten.