Congratsss PhD Christine Basta 👏🏼👏🏼👏🏼

Appreciation post to this people who were always there to support the PhD journey, I was glad to have such an amazing team mt@upc. Missing in the photo Magdalena Biesialska @bardia_rafieian Casimiro Carrino and my amazing supervisor Marta R. Costa-jussa.

If you are in Madrid, do not miss this expo from the BSC-CNS that showcases our work on sign language translation with deep learning. Lead by Dr Amanda Duarte, today run by Laia Tarrés & Gerard I. Gállego, and happily co-advised with Prof. Jordi TORRES.AI.

(1/3) Until now, AI translation has focused mainly on written languages. Universal Speech Translator (UST) is the 1st AI-powered speech-to-speech translation system for a primarily oral language, translating Hokkien, one of many primarily spoken languages. bit.ly/3CJP3ew

DLBCN 2022 will be awesome: we have received 41 speaker & 52 poster proposals. These figures confirm the high potential of Barcelona as a deep learning hub. Notification of acceptance by November 5. (Image from tweet by Stability AI + Hugging Face)

It’s been a while since we started some random experiments with the Perceiver for Direct ST. The result is a method (DLA) that allows to adapt the model's computational cost independently of input sequence length. 🚀💬 Paper: arxiv.org/abs/2210.16264 Code: github.com/mt-upc/s2t-per…

Are we fully utilizing our Speech Translation data?🤔 In our latest work we combat data scarcity in ST by proposing SegAugment, a segmentation-based data augmentation method that consistently increases translation quality by generating alternative versions of an ST corpus.

Would like to learn about the latest advances in #ArtificialIntelligence? A new edition of our postgraduate course starts in February, and there are still places available! The information session will be held next Wednesday, January 18th, at 6:30pm.

Would you like to join the MTG in Barcelona for your PhD? Now we are opening 4 funded PhD positions starting in October 2023. Deadline for applications: March 19th. Find here all the info about topics, supervisors, and requirements: upf.edu/web/mtg/news/-…

Very proud of this achievement and honored to be part of the team! Excited to share our #ACL2023 paper, “Explaining How Transformers Use Context to Build Predictions”. It has been an amazing journey diving deep into the Transformer! ✨

Presenting mt@upc's contribution to #IWSLT2023. This year, we're leveraging large speech & text models, and optimizing them for ST fine-tuning through Optimal Transport. Want to know more? Check the paper! 👇

Inseq 🐛 is a new toolkit built on top of the Hugging Face ecosystem to enable easy and reproducible interpretability analyses of generative LMs! Nils Feldhus and I will present Inseq today at #ACL2023NLP Demo Session 5 (16:15-17:30). Come say hi and get a sticker! 😄

📣 Career Update Alert! Just wrapped up a fantastic internship with Apple’s Global Siri team. Now, I'm gearing up for another internship, this time with Amazon in Aachen, Germany. Feeling incredibly grateful for these opportunities and super excited for what the future holds!

> next token prediction cannot lead to learning a world model. It's dumb. It's "just" stats. Next token prediction:

Check out our new work led by Lukas Edman on comparing subword and byte-level models for machine translation, with a nice interpretability analysis of character-level language generation!

Happy to share ZeroSwot: We present a method for Zero-shot Speech Translation, that not only surpasses previous zero-shot models, but also supervised ones, setting new SOTA in MuST-C and outperforming the original version of SeamlessM4T in CoVoST. arxiv.org/abs/2402.10422

📣 Exciting update! Moving from Amazon to Dolby Labs for my next internship adventure. Grateful for the invaluable experience gained and eager to embark on this new chapter. Huge thanks to my exceptional mentors and the supportive team! 🚀🎉