nusretozates.bsky.social
ML/Ops Engineer | MS at Koc University studying Computer Vision in Pathology | Working on NLP in the industry | Google Cloud Certified - Professional ML Engineer
44 posts
121 followers
1,789 following
Regular Contributor
Active Commenter
comment in response to
post
Some people came and started to use Twitter again in a very short time 🥲
comment in response to
post
I'm seeing those new optimizers and thinking about if I can use them with a small batch size (e.g 4 or 16) for my image segmentation tasks. What do you think?
comment in response to
post
Totally irrelevant but I just realized you are working at Riot and working on... LLMs? I'm really curious right now 😂
comment in response to
post
I have some questions:
- Can we fine-tune a model with registers and get same results?
- Can we do that with only last x layers?
- Given a trained dinov2 wo registers, would adding reg token help for my downstream tasks like segmentation?
comment in response to
post
arxiv.org/abs/2309.16588
comment in response to
post
comment in response to
post
I'm very open for ideas/tips/hints btw. :)
comment in response to
post
Yeap found it 😄
comment in response to
post
I think I saw a post about how it switch to Chinese when it "really" needs to thinks. And people were referring to Karpathy's tweet below
comment in response to
post
Additionally, I don't think you can be a senior in the AI field without theory knowledge nor one of the people who develop the technology of the future
comment in response to
post
I see thank you so much for the answer 🙏
comment in response to
post
For example, the second sentence in the introduction part is generally the first thing that thought in computer vision with deep learning classes. I feel like this is a common knowledge now and doesn't need any citation but seems like I'm wrong, right?
comment in response to
post
Yeah that is right 🤦♂️ It's nice to know thanks a lot for the answer!
comment in response to
post
I'm getting similar questions from university students about generative AI as it makes easier to clone voices, creating fake images etc. and as far as I see (not that much though) researchers mostly tries/prefers to ignore this kind of questions and push the quality higher
comment in response to
post
Or I can change the question to "Is number of citations a quality metric?" I feel like people force themself to add citations more than necessary
comment in response to
post
New question for me: What would I see if I make the last layer trainable, find a use of the CLS token for segmentation task and look at the attention again
comment in response to
post
The model's image representation is very strong for segmentation but attention visualization of the CLS token cannot show you the objects in the image like DINO V2 trained on ImageNet. Funny fact: This model's CLS token also attends to the bird if you give an image of a bird 😂
comment in response to
post
She is right though and I'm super scared about finding a nice thesis topic with a nice technical contribution
comment in response to
post
I'm working on AI in Pathology for my MSc, can you also add me 🙋♂️
comment in response to
post
Having a friend at the target company >>> lots of other things
At minimum, it gives a %99 chance that a human will read your CV and interview with you