Katherine Tian
@kattian_
Followers
864
Following
558
Media
3
Statuses
131
cs/stat @harvard , working on calibration & factuality of LLMs, prev @GoogleAI tensorflow, golden state @warriors fan
Bay Area, CA
Joined January 2018
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Georgia
• 349198 Tweets
Texas
• 283706 Tweets
Yankees
• 240299 Tweets
Never Trump
• 161662 Tweets
Arnold Palmer
• 137238 Tweets
Presiden
• 100902 Tweets
World Series
• 91818 Tweets
WANG YIBO IN THE GT FINALS
• 67199 Tweets
Juan Soto
• 61803 Tweets
期日前投票
• 61213 Tweets
CHENLE VICTORIOUS FIRST THROW
• 54629 Tweets
Stanton
• 44933 Tweets
GALA EN EL AUDITORIO
• 40178 Tweets
#DOMELIVE_Atlantis
• 33464 Tweets
Kirby
• 29215 Tweets
Rayados
• 25885 Tweets
#菊花賞
• 23703 Tweets
ZETA
• 23658 Tweets
#刀剣乱舞ONLINEもうすぐ十周年
• 22068 Tweets
#かぼちゃ大作戦シルエットクイズ
• 22021 Tweets
Ewers
• 19277 Tweets
Tigres
• 18822 Tweets
gerard
• 17564 Tweets
ダノンデサイル
• 16872 Tweets
アーバンシック
• 13840 Tweets
ねこあつめ2
• 11378 Tweets
コスモキュランダ
• 10216 Tweets
Pinned Tweet
Thanks Paige! Great working on the
@TensorFlow
team
Announcing a cool new feature in
@TensorFlow
: differentiable map ops! 🙌✨
If you've ever wanted to store and train embeddings in an embedding model, this should make the process *much* simpler. 😄
Amazing work from our TF intern,
@kattian_
(
@Harvard
)!
2
49
199
1
3
44
Turns out we can directly train language models to hallucinate less without any human annotation -- for around 50% error reduction compared to RLHF!! Check out our paper for the approaches and full results 😃
Fine-tuning Language Models for Factuality
paper page:
The fluency and creativity of large pre-trained language models (LLMs) have led to their widespread use, sometimes even as a replacement for traditional search engines. Yet language models are prone
5
48
255
12
36
247
How can you tell how reliable an LLM output is?
Let’s talk about calibration—calibrated confidences give an accurate probability of whether an LLM output is correct AKA how reliable it is
Our EMNLP paper shows we can get calibrated confidences from RLHF LLMs by *just asking* 🧵
5
14
99
Check it out!! It was great working with Jaehwan + the team and thanks to Pranav for advising 😊
1/ Introducing X-REM, our new AI method to generate radiology reports from chest X-ray images.
Our method uses a multimodal language-image model to capture fine-grained interaction between text and image for higher accuracy.
#AI
#radiology
#MIDL2023
🧵
4
36
155
1
2
15
Check out our paper for the full prompts, results, and takeaways across GPT-3.5, GPT-4, Claude, Claude-2 and Llama2-7B-Chat!
3
3
14
Check out the JAX transformer! Exciting to see how JAX did in MLPerf
Very excited to see the MLPerf 0.7 results released today, where Google TPUs set records in six of the eight benchmarks!
We need bigger benchmarks, because we can now train the ResNet-50, BERT, Transformer, & SSD benchmarks each in under 30 seconds.
20
279
1K
0
2
8
Had a great time presenting this interesting NLP interpretability paper for Harvard Undergrad ML Group's journal club today! Slides:
Join the discussion now! My PhD student
@danish037
& CMU alum
@im_mansigupta
are discussing our work on Learning to Deceive with Attention-Based Explanations
0
1
15
1
0
6
Excited!!
Course Launch📢 AI Research Experiences
🗓️ Over the past 5 years, I've been training students to get their start in applied AI engineering and research.
Over the next 4 months, as I launch CS197 at Harvard, we will be making our notes publicly available:
6
55
245
1
0
5
This project was done with the incredible
@ericmitchellai
and team
@AllanZhou17
@archit_sharma97
@rm_rafailov
@HuaxiuYaoML
, supervised by
@chelseabfinn
@chrmanning
😃
0
0
6
great perspective worth hearing! thanks for sharing, av
“On Motivational Debt, Shame Avoidance, Impostor Syndrome and Living Well”
Alternative title: Av does SURPRISE THERAPY at a weekly group meeting
Response was very positive so decided to record & share!
Consider discussing in your group meeting :-)
7
5
32
0
1
5
Can we avoid log probs and *just ask* the model to verbalize its confidence as text? Yes! Verbalization can cut expected calibration error in half for models including GPT-4, Claude-2, Llama2-70B-Chat
1
0
4
However, evidence suggests that RLHF destroys the calibration of LM’s output distribution. See the GPT-4 technical report and our confirmation on Llama2-70B
1
0
6
👇👇Raaz is a great mentor and super fun to work with!
PhD Applicants: Consider the vibrant research
@cornell_tech
, where
@cornell
meets New York City!
Interested in statistical problems in causal inference & reinforcement learning for personalized decision-making & healthcare? Mention me!
RTs appreciated!
0
47
168
0
0
4
Other observations:
- Asking the model for multiple guesses and probabilities (inspired by human psychology) reduces overconfidence
- Surprisingly, despite LLM’s challenges with numbers, asking for verbalization with numbers worked as well or better than with phrases
1
1
4
@nibnalin
@amirbolous
that’s a reasonable hypothesis but it doesn’t seem like that’s the full story - evan chen reviewed a subset of proofs and verified they’re not all just coordinate bashing
0
0
5
How? We just ask the model for its best guess and probability that it is correct
1
0
2
Typically, you can get pretty well-calibrated confidences just from the model output distribution log probabilities or sampling from the output distribution
1
0
2
@kundan_official
Good question it could be! For intuition, the max frequency ignores the randomness of the model's sample at that token and rewards good prefixes; Also both options should be correlated since the max frequency is likely to be chosen. Will probably check empirically sometime
1
0
2
1
0
1
1
0
1
0
0
1
@valor_zhang
jump rope and cartwheels are more fun than running (also spikeball if you can get 3 others to regularly play with u)
0
0
1
1
0
1