Andrea Banino
@AndreaBanino
Followers
2K
Following
1K
Media
5
Statuses
255
I work at @DeepMind, I’m a machine learning researcher working on artificial general intelligence. I also want to understand how our brain works.
London, England
Joined August 2010
🚀 Excited to share our latest work @GoogleDeepMind: "Synth^2: Boosting Visual-Language Models with Synthetic Captions and Image Embeddings" 🧵 #VLM #Synthetic #AI
https://t.co/ID9LLebo6l
arxiv.org
The creation of high-quality human-labeled image-caption datasets presents a significant bottleneck in the development of Visual-Language Models (VLMs). In this work, we investigate an approach...
3
12
52
Work done by amazing people at GDM and led by @sahandsharif
0
0
0
8/ 🎓 Key Contributions: • Fully synthetic high-quality text-image pairs for VLM training • Efficient image embedding generation • Controlled study for fair evaluation • Significant performance gains in scene description, QA, and external knowledge QA
1
0
1
7/ 🔬 By pre-training both the text-to-image model and VLM on the same dataset, we isolate the true benefits of synthetic images, ensuring improvements are due to our method, not pre-existing knowledge.
1
0
0
6/ 🌍 Our work opens up promising avenues for developing self-improving multimodal models, addressing data scarcity, high curation costs, and noise in traditional datasets.
1
0
0
5/ ⚡ Synthesizing images in the embedding space is 25% faster than in the pixel space, reducing memory overhead and resource consumption without compromising data quality.
1
0
0
4/ 🔄 We found that semantic diversity and balance in captions are crucial for better downstream performance. Our analysis provides new insights into optimizing synthetic data for VLM training.
1
0
0
3/ 📊 Extensive experiments show our VLM, finetuned on synthetic data, performs comparably to models trained on human-annotated data, but with significantly less data! This demonstrates the power and efficiency of our synthetic approach.
1
0
0
2/ 🖼️ Our method employs a pretrained text-to-image model to generate image embeddings from LLM-generated captions. This approach expands beyond the original dataset, creating novel compositions that enrich VLM training data.
1
0
0
1/ 🔍 Tackling the bottleneck of high-quality human-labeled datasets for Visual-Language Models (VLMs), we propose a novel approach using Large Language Models (LLMs) and image generation models to create synthetic image-text pairs.
1
0
0
🚀 Excited to share our latest work: "Synth2: Boosting Visual-Language Models with Synthetic Captions and Image Embeddings" https://t.co/wynFKygi8f🧵
#GenAI #VLM
arxiv.org
The creation of high-quality human-labeled image-caption datasets presents a significant bottleneck in the development of Visual-Language Models (VLMs). In this work, we investigate an approach...
2
1
4
Google announces Synth^2 Boosting Visual-Language Models with Synthetic Captions and Image Embeddings The creation of high-quality human-labeled image-caption datasets presents a significant bottleneck in the development of Visual-Language Models (VLMs). We propose a novel
6
80
347
That’s on again!sign-up if you want to have the opportunity to learn the most recent breakthroughs in ML!bonus: this time is by the sea 🌊🏄♂️
We are excited to announce the 3rd edition of the Mediterranean Machine Learning (M2L) summer school in August 2023! This year, the school will take place at @actgreece in Thessaloniki, Greece. Apply at
0
0
1
After 2.5+ years of pandemic, I will attend my first in-person meeting in September! https://t.co/Eejm2IHZyG So glad to meet people again! :) Thanks @AndreaBanino and the organizers for the invitation.
m2lschool.org
Mediterranean Machine Learning Summer School University of Split Faculty of Electrical Engineering, Mechanical Engineering and Naval Architecture Split, Croatia 8-12 September 2025
0
3
22
Announcing the 1st Dynamic Neural Networks (DyNN) workshop, a hybrid event @icmlconf 2022! 👇 We hope DyNN can promote discussions on innovative neural networks that can deal with dynamic computations. Want to learn more?
dynn-icml2022.github.io
Friday, July 22 - 2022 International Conference on Machine Learning - Baltimore, MD
0
17
58
I’m very happy to present this work that has been accepted as spotlight paper at ICLR. Train large models in RL will be critical to learn better value functions, we are proposing a way to make this learning more efficient in terms of data. Check it out!
CoBERL, a simple and scalable method to improve data efficiency in a variety of RL environments (Atari, DmControl, DmLab). Lean more today https://t.co/FCtZn8u2Gx
#ICLR2022
0
1
2
New paper! Language and large foundation models come together to drive semantically meaningful exploration. This idea helps RL agents learn faster in 3D environments, even when language annotations are unavailable ( https://t.co/ez0PGuwFXC) Read on 🔎⬇️
7
84
445
We are pleased to announce the first set of speakers. Thanks to @ChrisBishopMSFT, @MihaelaVDS, @aprorok, @_aidan_clark_, @elluba, @MihaelaCRosca, @yayitsamyzhang, @luisa_zintgraf. If you haven’t done it yet, apply at: https://t.co/7Bgn5XMhoQ
#AI #ML #RL #M2Lschool
docs.google.com
Important information: =================== 1. Please READ ALL THE DESCRIPTIONS carefully. Failure to read and/or follow the instructions will result in the application being desk rejected. 2. You...
0
10
27
Last few days to sign up to the M2L summer school!Remember, this year it will be free for all students thanks to our amazing sponsors! Apply at https://t.co/ZmwcJkyxG6!
#AI #ML #RL #deeplearning #machinelearning #M2Lschool 1/2
1
8
19
Two weeks left to apply to the M2L school! Lots of lectures, tutorials, great speakers, and much more! Apply here: https://t.co/7Bgn5XMhoQ All info at https://t.co/pOybcCNyaV
0
8
19