Oğuzhan Ercan

Representation Learning @tubitak

[Oğuzhan Ercan @oguzhannercan]

@peer_rich person of interest

https://t.co/raVvj2dRKN

[Oğuzhan Ercan @oguzhannercan]

@alexanderkr low fps PID

[Oğuzhan Ercan @oguzhannercan]

@OpenAI @giffmana your time has come

[Oğuzhan Ercan @oguzhannercan]

@dr_cintas disk killer

[Oğuzhan Ercan @oguzhannercan]

@gdb Quick question: have you changed the weighting toward recent conversation context? Right now, ChatGPT seems overly biased toward referencing older recent chats in its answers.

[Oğuzhan Ercan @oguzhannercan]

@SwayStar123 I think model performs well with this size. I added and image. At the feature space, it behaves almost same with SVG (2510.15301) which uses dinov3 for feature alignment. Congrats.

[Oğuzhan Ercan @oguzhannercan]

@massiviola01 Totally agree, but it works :) from my exps I can confirm that moes improve out of distribution accuracy at img classification.

For decades, all of novelty was coming from computer vision due to military funds but right now money goes to llms so novelties are.

[Oğuzhan Ercan @oguzhannercan]

@giffmana I did quick review on paper, so i must missed that part, thanks. The reason why I have focused this sample was it is actually correctly labeled at 2012 IN. Trying to push boundaries to %100, i think i should find a way to re-label misclassified samples.

[Oğuzhan Ercan @oguzhannercan]

Are we done with labeling imagenet @giffmana? Best of i can see (if i did not make silly mistake), still there are labeling errors at reassessed-imagenet validation set. Maybe both paper_towel and bakery could be accepted as label, but only paper_towel seems incorrect.

[Oğuzhan Ercan @oguzhannercan]

@ai_impact_rank from unidecode import unidecode
unidecode(entity) # gökberk ->gokberk

[Oğuzhan Ercan @oguzhannercan]

@ai_impact_rank Here there is a bug. The entity is listed twice with different scores.

[Oğuzhan Ercan @oguzhannercan]

@Hesamation @1Punch_Girl Why would I memorize the number of heads? Explaning it with warp, tile or data transaction shows the understanding of how things happen at low level. Other is memorization.

[Oğuzhan Ercan @oguzhannercan]

@svpino @that_coder There is a joke

[Oğuzhan Ercan @oguzhannercan]

@giffmana @YiTayML Just programmer

[Oğuzhan Ercan @oguzhannercan]

@loganthorneloe companies looking for Goodfellow (generative adversarial nets: 2014)

[Oğuzhan Ercan @oguzhannercan]

@cloneofsimo Still betting on yann

[Oğuzhan Ercan @oguzhannercan]

@attentionmech Cs 492 kaist yt

[Oğuzhan Ercan @oguzhannercan]

@attentionmech Actually not scary (after spending 2 years on that)