alvin!
@alvanlii
Followers
201
Following
382
Media
18
Statuses
52
uw, building tts models for a low-resource language
Toronto, Ontario
Joined March 2020
if you are interested in our future meetups, subscribe to our event calendar! We usually do Saturdays at 12 .
luma.com
View and subscribe to events from 🌳Papers in the Park on Luma. At Papers in the Park we unpack one AI research paper together—no slides, minimal jargon, maximum curiosity—while the breeze turns...
1
0
2
During the event, we looked at the models leading up to v3, including SimCLR and iBOT. DINOv3 is about scaling the encoder model and tackling degraded feature maps using Gram matrices. It also talks about aligning image features with text models.
arxiv.org
Self-supervised learning holds the promise of eliminating the need for manual data annotation, enabling models to scale effortlessly to massive datasets and larger architectures. By not being...
1
0
1
when I was in university, DINO is one of my fav papers and it was great to revisit it. (screenshot from UW data science club IG when I did a takeover 3 years ago)
1
1
4
Week 11 of Papers in the Park 🌳 with @Cohere_Labs @asusevski. We reached double digits! So grateful for everyone who came and continue to support our events. This week we covered DINOv3
1
6
46
@Cohere_Labs @asusevski if you are interested in our future events, subscribe to our Luma calendar! .
luma.com
View and subscribe to events from 🌳Papers in the Park on Luma. At Papers in the Park we unpack one AI research paper together—no slides, minimal jargon, maximum curiosity—while the breeze turns...
0
0
5
@Cohere_Labs @asusevski second paper was Persona Vectors. This Anthropic paper talks about how persona vectors can be used for monitoring + reducing persona shifts, and identifying data that can contribute to emergent misalignment. I said "evil" wayy too many times during this.
arxiv.org
Large language models interact with users through a simulated 'Assistant' persona. While the Assistant is typically trained to be helpful, harmless, and honest, it sometimes deviates from these...
1
0
3
@Cohere_Labs @asusevski first paper was GRAINS. GRAINS is about using integrated gradients to find tokens that contributes to positive/negative outputs, then using them to create steering vectors to alleviate hallucinations in LLMs and VLMs.
arxiv.org
Inference-time steering methods offer a lightweight alternative to fine-tuning large language models (LLMs) and vision-language models (VLMs) by modifying internal activations at test time without...
1
0
4
Toronto's Papers in the Park Week 9.with @Cohere_Labs @asusevski. It's great to see new and returning folks every week, people from different backgrounds gathered to talk about ML stuff. This week we looked at 2 papers on model steering:
4
6
46
STITCH:Â STITCH is about how we can leverage the difference between perceived audio duration and token generation time to reason in between.
arxiv.org
Spoken Language Models (SLMs) are designed to take speech inputs and produce spoken responses. However, current SLMs lack the ability to perform an internal, unspoken thinking process before...
1
0
1
Your LLM Knows the Future: Uncovering Its Multi-Token Prediction Potential:Â This paper talks about how an LLM can be adapted to predict multiple tokens at the same time, as opposed to one token at a time.
1
0
1
Thanks everyone for coming to the 7th (!!!) Papers in the Park in Toronto. We covered 2 papers today, one from Microsoft about Speech LMs and one from Apple about Multi-Token generation.
3
4
19
Papers in the Park is back again this Saturday!. Join us and dive into research papers on a chill summer day. Everyone is welcome! 🌳. @Cohere_Labs
1
5
30
RT @Cohere_Labs: 📣 Toronto Researchers: Papers in the Park is back this Saturday!. Join us as we unpack research papers together— It's a lo….
luma.com
View and subscribe to events from 🌳Papers in the Park on Luma. At Papers in the Park we unpack one AI research paper together—no slides, minimal jargon, maximum curiosity—while the breeze turns...
0
7
0
did it as part of a hackathon 2-ish weeks ago with @asusevski @_aandyw , it's like a thing where you choose two MPPs and a topic and they debate on it
1
2
10
