
Assistant Professor in Medical Image Analysis @TUeindhoven. All expressed opinions are my own. The "weird" language I sometimes tweet/retweet is Macedonian.
Eindhoven, Nederland
Joined March 2008
- Tweets5.5K
- Following901
- Followers1.2K
- Likes2.8K
Mitko Veta retweeted
GenAI does not stop, so it is time for a new blog post. Since LLMs are everywhere, I decided to take a look at them, my curious readers. And a bonus: An implementation of a teenyGPT 🤖✨
📄Post: jmtomczak.github.io/blog/20/…
🖥️Code: github.com/jmtomczak/intro_d…
Mitko Veta retweeted
Don't forget about applying to Generative Modeling Summer School on 24 - 28 Jun 2024 (Eindhoven, NL): d6jx7.app.goo.gl/G6DioFUGBpV…
Mitko Veta retweeted
We are excited to announce PathChat - a vision-language AI assistant for #Pathology that can analyze histology images and answer diverse pathology-related queries.
Co-led by our superstars @MYLu97 @chenbowen118 @DFKW_MD
Preprint: arxiv.org/abs/2312.07814
Demo below,
Mitko Veta retweeted
The challenge report of the MIDOG #MICCAI #challenge has just been released as a preprint! Thanks to all contributors who made this possible :-)
Find a brief summary of the insights in this blog post: deepmicroscopy.org/insights-…
Though our Al model shows a significant discriminative value, it didn't improve over established predictors. Indicates that the value of CT imaging for predicting response is likely limited. Brings important nuance to positive findings in this field. #oncology #AI #radiomics
Link to the preprint: medrxiv.org/content/10.1101/…
85
Mitko Veta retweeted
We just released the MIDOG++ multi-domain mitosis dataset! With 504 cases, 12k mitoses (3 expert consensus) and 7 tumor domains it is the largest to date. The main insights from the paper are in this 🧵
deepmicroscopy.org/midog-the…
Mitko Veta retweeted
Attention all #ECDP2023 attendees: #Tweetup tomorrow at 13:45 at registration, spread the word! #MOTTIRL (meet-on-Twitter-then-in-real-life), #digitalpathology #pathtwitter #pathology
Mitko Veta retweeted
AI-generated QR codes using ControlNet are insane.
This is going to be increasingly common in ads in the near future.
These examples blew my mind (try scanning them):
1. Ancient Village
Mitko Veta retweeted
Clock drawing. It remains one of the easiest and cheapest ways to identify patients with mental/behavioral disorders who require further neurological & neuropsychological workup
Source: agsjournals.onlinelibrary.wi…
📢 PhD Opportunity! Join our interdisciplinary team at @TUeEAISI with @Ilsebyl and Liesbeth Janssen in a groundbreaking journey into AI, Physics & Oncology! Contribute to a cutting-edge project to model & predict tumor cell dynamics with AI. 🧬💻🔬
jobs.tue.nl/en/vacancy/phd-o…
Mitko Veta retweeted
Here's a World War 2 movie trailer I made this weekend, completely generated using AI text-to-video!
Script written by GPT-4, Michael Caine voiceover with @elevenlabs, and video AI generated with @runwayml Gen-2 😮
We still have a ways to go but the potential is so exciting!
The Dutch government, which is currently led by a ***liberal*** party, has decided to increase traffic fines. The decision was made after last year's budget fell short of expected revenue from fines 🤦♂️🤦♂️🤦♂️
nu.nl/economie/6261417/verke…
2
388
Mitko Veta retweeted
🇧🇪 Belgian customs crushed over 2,000 cans of Miller High Life after Champagne producers complained about the beer's slogan: "The Champagne of Beers". euronews.com/culture/2023/04…
Just kicked off our hybrid workshop on AI for spectroscopy as part of the ITEA Spectralligence project. Features talks by 10 research groups and companies. @JohanSvdBrink
Mitko Veta retweeted
Here Are 14 Characters Who Would Absolutely Pay for Twitter Blue dlvr.it/SmVTwB
Mitko Veta retweeted
Love it 👏 - much fertile soil for indie games populated with AutoGPTs, puts "Open World" to shame. Simulates a society with agents, emergent social dynamics.
Paper: arxiv.org/abs/2304.03442
Demo: reverie.herokuapp.com/arXiv_…
Authors: @joon_s_pk @msbernst @percyliang @merrierm et al.
Mitko Veta retweeted
This is a baby GPT with two tokens 0/1 and context length of 3, viewing it as a finite state markov chain. It was trained on the sequence "111101111011110" for 50 iterations. The parameters and the architecture of the Transformer modifies the probabilities on the arrows.
E.g. we can see that:
- state 101 deterministically transitions to 011 in the training data, so the probability of that transition becomes higher (79%). Not near 100% because we only did 50 steps of optimization.
- state 111 goes to 111 and 110 with 50% probability each, which the model almost learns (45%, 55%).
- states like 000 are never encountered during training, but have relatively sharp transition probabilities, e.g. 73% of going to 001. This is a consequence of inductive biases in the Transformer. One might imagine wanting this to be 50%, except in a real deployment almost every input sequence is unique, not present in the training data verbatim.
Not really sure where I was going with this :D, I think it's interesting to train/study tiny GPTs because it becomes tractable to visualize and get an intuitive sense of the entire dynamical system. Play with here: colab.research.google.com/dr…




















