Soumya Ghosh
@soumy_aghosh
Followers
145
Following
233
Media
14
Statuses
79
Director of ML at Merck. Prev: Statistical Machine Learning @IBMResearch and @MITIBMLab and @BrownCSDept
Boston / Cambridge, MA
Joined May 2019
A nice writeup in MIT news about our work on calibrating llms.
Thermometer, a method for calibrating a large language model from MIT and @MITIBMLab researchers, could help pinpoint situations where a LLM is overconfident and enable the model to produce better-calibrated responses on tasks it has not seen before.
0
0
9
RT @maohaos2: Thanks @soumy_aghosh for the nice introduction! We will be presenting our work at the first poster session on Tuesday in Hal….
0
3
0
0
0
1
Leads to a simple yet effective procedure. Preserves accuracy of the llm when greedy decoding, has an interesting variational interpretation. Work led by the amazing @maohaos2 (follow him). Paper: Code (coming soon):
github.com
Contribute to maohaos2/Thermometer development by creating an account on GitHub.
1
0
1
Our work "Thermometer" at #ICML2024 improves calibration properties of llms without multiple forward passes or labeled calibration data. We build on temperature scaling and learn a mapping from groups of prompts to temperature allowing us to predict the temperature for new groups
1
3
10
RT @krvarshney: Trans-LoRA is really excellent work by Runqiang Wang, @soumy_aghosh, @neurobongo, et al. that can really help the community….
0
3
0
RT @leokarlin: Thanks for the highlight @_akhaliq!.We offer a simple and nearly-data-free way to move (large quantities) of custom PEFT mod….
0
6
0
@leokarlin With as few as 5 fine-tuning instances, we are able to transfer across models with virtually no loss (and often improved) in performance.
0
0
0
Excited about this new work led by Ray Wang and @leokarlin. We transfer PEFT modules across LLMs by using a synthetic data generator trained to approximate the data generating process of the observed fine-tuning data.
Trans-LoRA. towards data-free Transferable Parameter Efficient Finetuning. Low-rank adapters (LoRA) and their variants are popular parameter-efficient fine-tuning (PEFT) techniques that closely match full model fine-tune performance while requiring only
1
0
2
It was a lot of fun attending Siam UQ last week and talking about BNNs with horseshoe priors and the “are you using test log-likelihood correctly?” work at the sparsity in Bayesian UQ mini-symposium.
0
0
6
Now at TMLR! We explore the well known yet often forgotten fact that test log-likelihood measures closeness of a predictive distribution to the true data generating process in a certain KL sense. It might not convey anything useful about the quality of a posterior approximation.
Are you using test log-likelihood correctly?. Sameer Deshpande, Soumya Ghosh, Tin D. Nguyen, Tamara Broderick. Action editor: Michael Gutmann. #likelihoods #likelihood #comparisons.
1
2
18
RT @justindubinmd: So I have never watched Bluey but a lot of my friends who have kids do. Apparently Season 2 Episode 9 titled “Sleepyti….
0
2K
0
RT @variational_i: The LLM Alignment team at IBM Research is looking for a talented PhD student for a summer internship at the MIT-IBM Wats….
0
10
0
RT @krvarshney: New URLs for the @IBMResearch websites for the Trust 360 toolkits:. * AI Fairness 360 * AI Explaina….
0
11
0
To be fair, the other reviews were critical but helpful. Will definitely help improve the paper.
1
0
0
R2 makes a bunch of factually incorrect assertions, grossly misinterprets experiments and plots, misses important details. The rebuttal painstakingly points out the errors and misunderstandings. R2 promptly responds: 🙃
1
0
2
RT @IBMWatson: Our next-generation enterprise studio for #AI builders: IBM brings together traditional machine lear….
0
27
0