🚨Instead of collecting costly datasets for tabular prediction, could we use natural language instructions?💡
In our paper "TABLET: Learning From Instructions For Tabular Data" with
@sameer_
we evaluate how close we are to this goal and the key limitations of current LLMs
Imagine having conversations🗣️ with ML models like you would with a colleague💡
In our paper "Rethinking Explainability as a Dialogue: A Practitioner's Perspective" with
@hima_lakkaraju
@sameer_
Yuxin Chen and
@ChenhaoTan
, we evaluate this possibility!🧵
Explaining ML models can be a difficult process.
What if anyone could understand ML models using accessible conversations?
We built a system, TalkToModel, to do this!🎉
Paper:
Code & Demo:
1/9
I'll be at
#NeurIPS2023
next week presenting our work "Post Hoc Explanations of Language Models Can Improve Language Models"
🎉reach out if you want to catch up!🎉
Explaining ML models can be a difficult process.
What if anyone could understand ML models using accessible conversations?
We built a system, TalkToModel, to do this!🎉
Paper:
Code & Demo:
1/9
Interested in our group’s recent work in
#MachineLearning
explainability and where we think things are heading?
I gave a seminar
@UCIbrenICS
yesterday about this. Check out the recording!
Explaining ML models can be a difficult process.
What if anyone could understand ML models using accessible conversations?
We built a system, TalkToModel, to do this!🎉
Paper:
Code & Demo:
1/9
In non-COVID news, I'm going to start posting about a paper every few weeks. I read an interesting paper about efficient approximations of Shapley values recently () that provides two neat methods to do this based on data structure. 1/n
📢Come check out our NeurIPS paper later today🗓️ "Reliable Post hoc Explanations: Modeling Uncertainty in Explainability"
We introduce a method to generate more robust local explanations using uncertainty
Poster:
Date: 9 Dec 4:30pm PST — 6pm PST🧵👇
No book should be on this list, but EVERYWHERE BABIES??? It is a board book celebrating babies being babies while adults love, feed, and care for them.
Presenting two papers this year
@NeurIPSConf
!
The first paper, "Reliable Post hoc Explanations: Modeling Uncertainty in Explainability" concerns generating robust local explanations
Paper:
Poster:
Date: Dec 9 4:30pm — 6 PST
Ever wondered when you shouldn't use a fair ML model? Pleased to share our new paper in
#FAT2020
"Fairness Warnings & Fair-MAML: Learning Fairly from Minimal Data " (w/
@kdphd
, twitterless Emile) where we investigate such questions.
How reliable are fairness metrics with limited data?
*They're not* but we can do better using unlabeled data + Bayesian interference.
Highlighting work from colleagues
@ji_disi
, Padhraic Smyth, and Mark Steyvers from NeurIPS.
How to assess? [1/10]👇
Often, the choice of superpixel generating function & hyperparameters is ignored for local image explanations
This choice has significant effects on explanation faithfulness
Let's sweep hyp. params for different superpixel algs & run LIME explanations for 50 imagenet images
Exciting next couple weeks presenting work! I'll be at FAT* this week to talk about our recent paper "Fairness Warnings & Fair-MAML: Learning Fairly from Minimal Data" ()
Want to know how adversaries can game explainability techniques? Our latest research - "How can we fool LIME and SHAP? Adversarial Attacks on Explanation Methods" has answers: . Joint work with the awesome team:
@dylanslack20
, Sophie, Emily,
@sameer_
#acl2020nlp
If you are going to grad school, apply to Sameer Singh's group (
@sameer_
). It is hard to find problems to work on, and his work consistently targets new and impactful areas. We are always kicking ourselves two years later rereading his papers.
How to find model errors beyond those available in the data?
We propose methods to automatically generate high-level "model-bugs" in image classifiers.
This preprint includes some of my summer work
@awscloud
with
@kkenthapadi
and Nathalie Rauschmayr.
anyways, if you're looking to read other ~edgy~ titles like "sweet stories for babies", "bear in the air", or "puppies, puppies, puppies" check out her stuff at
You might know that MSFT has released a 154-page paper () on
#OpenAI
#GPT4
, but do you know they also commented out many parts from the original version?
🧵: A thread of hidden information from their latex source code
[1/n]
🌇Overall, we hope our work serves as a good starting place for engineers & researchers to design interactive, natural language dialogue systems for explainability that better serve users’ needs.
📑
📢Come check out our NeurIPS paper later today🗓️ "Reliable Post hoc Explanations: Modeling Uncertainty in Explainability"
We introduce a method to generate more robust local explanations using uncertainty
Poster:
Date: 9 Dec 4:30pm PST — 6pm PST🧵👇
It's been pretty fun to explore the sample space of
#stablediffusion
& really impressed by how expressive it is, especially considering the *relatively* smaller model size
This one is from the prompt "a lithograph of a butterfly"
Next week, I'll be at AIES presenting a paper on post-hoc interpretation attack techniques "Fooling LIME and SHAP: Adversarial Attacks on Post hoc Explanation Methods" ().
The second paper, "Counterfactual Explanations Can Be Manipulated" addresses how to adversaries might manipulate gradient-based counterfactual explanations and how this can be prevented.
Paper:
Poster:
This Thursday,
@dylanslack20
from UC Irvine will be talking to us about exposing shortcomings & improving the reliability of machine learning explanations. Catch it at 1-2pm PT this Thursday on Zoom!
Subscribe to
#ML
#AI
#medicine
#healthcare
pretty crazy to me my grandmother, a person who writes cute and loving children's books about bunny rabbits, dogs, and cats, now has a banned book in the US
Our system helps address a few critical difficulties with explanations
1/ Deciding which explanations to use
2/ Computing the explanations
3/ Interpreting the results
4/ Engaging with further questions beyond the original explanation
2/9
Two of our papers just got accepted for oral presentation at AAAI Conference on AI and Ethics (AIES):
1. Designing adversarial attacks on explanation techniques ()
2. How misleading explanations can be used to game user trust? ()
Both these papers are with Sophie Hilgard,
@hima_lakkaraju
, and
@sameer_
.
If you're interested in chatting about either of the papers, please reach out😀
26 people across 10 institutions 🤯🤯🤯
NLP models can learn artifacts in data instead of solving the actual task, leading to good performance #-wise but not really language understanding.
Check out this impressive & compelling project
@nlpmattg
led!
Evaluating NLP Models via Contrast Sets
New work that is a collaboration between 26 people at 10 institutions (!)
Trying to tag everyone at the top of the thread, here it goes:
Here’s the job link for joining SEAL:
If you have questions about the role feel free to DM me. I might not be able to get through all the pings but I’ll start reviewing all the applications next Friday.
I'll be at NeurIPS this week until Saturday, presenting two of our recent works on explaining ML models with conversations! Reach out if you want to chat😀
[Workshop Paper] TalkToModel: Explaining Machine Learning Models with Interactive Natural Language Conversations (joint w/
@dylanslack20
@SatyaXploringAI
@sameer_
) at TSRML workshop -- . More details in this thread [9/N]
We release the code for TalkToModel and link to a demo on a diabetes prediction task:
TalkToModel can be extended, modified, and adapted to your own models and datasets. We provide tutorials here:
8/9
@arrayslayer
@hima_lakkaraju
@sameer_
We look at model agnostic explanation methods -- those that don't exploit the specific structure of a model to perform explanations. TreeSHAP has access to full tree structure of the model so would be outside the scope of this attack.
First, we interviewed doctors👩⚕️, healthcare professionals⚕️, and policymakers👨💼 about their
1/experiences with existing explainability methods
2/needs & desires for future techniques
We learned A LOT about what practitioners want out of explainability in these conversations!
In spite of all its hype and the vitriol against it, the bad reviews and unreasonable rejects, despite the uncharitable takes by everyone on everything, this is SUCH an exciting time to do research in ML. So much fun seeing all the new work. Please DO tweet about it. Seriously.
🗣️ "I don't know anything about how correct the explanation is! How do you expect me to use it meaningfully? I constantly struggle with worrying about using an incorrect explanation and missing out on not using a correct explanation that is giving me more insights."
🚨However, we also observed several critical limitations of LLMs as well on these tasks☝️
1/ LLMs predict the same thing on logically inverted instructions, indicating unfaithfulness
2/ LLMs misclassify instances across many possible few-shot examples, indicating biases
Also, GitHub plug: . We tried to make these models pretty user friendly. It's fun to play around with data sets to see what you can get LIME/SHAP to explain!
Overall, we found,
1⃣ experts aren't satisfied w/ current explainability methods
2⃣ they want increased interaction to understand model behavior over one-off explanations
In our evaluation of TABLET, we found natural language instructions were quite helpful for LLMs in solving tasks solely from instructions (i.e., the zero-shot setting)😀
Overall, enjoyed reading this paper and recommend checking it out. I know there's some larger concerns floating around about the use of Shapley values to assign feature importance () and recommend reading these as well. 11\n
Super excited to share that I recently received an NSF grant to work on exposing vulnerabilities of post hoc explanation methods and enhancing their robustness. More details at
TalkToModel learns to parse user inputs for a new model and dataset into a programming language designed for model understanding.
It then executes these instructions, potentially comparing many explanations to ensure accuracy, and composes the results into a response.
3/9
@willieboag
Glad you thought it was interesting! Also shouts out to team Sophie, Emily,
@hima_lakkaraju
, and
@sameer_
. Code not linked to in this version of the paper but if you're interested:
📦We compile the tasks in TABLET from different sources:
1/ UCI datasets, such as credit, adult, churn
2/ Differential diagnosis (ddx) tasks, such as identifying whooping cough from patient symptoms
Explaining ML models can be a difficult process.
What if anyone could understand ML models using accessible conversations?
We built a system, TalkToModel, to do this!🎉
Paper:
Code & Demo:
1/9
tried this optimizer on a synthetic polynomial factorization task with a encoder/decoder transformer (e.g., 2*s*(26-7*s) ==> -14*s**2+52*s)
It's lagging a bit behind Adam/AdamW w.r.t. validation loss, but maybe I should be looking at perplexity/accuracy considering authors...
"Adan: Adaptive Nesterov Momentum Algorithm for Faster Optimizing Deep Models"
Another optimizer paper attempting to descend through a crowded valley to beat Adam. But...maybe this one actually does? [1/11]
@tallinzen
I've had some good interactions on gather town—that said there have been several frustrating technical issues (posters won't load on gathertown, login issues with underline)
Finally, we evaluate TalkToModel in human trials.
We compare health care worker and ML practitioner performance on several model understanding tasks using TalkToModel and a popular point-and-click dashboard as a baseline.
6/9
👉TABLET consists of a benchmark of tabular prediction tasks annotated with several different natural language instructions that vary in **complexity, phrasing, and source**
We evaluate how well models use natural language instructions for tabular prediction with TABLET!
📜We collect the natural language instructions from high-quality sources such as NIH, Merck Manual, and the national library of Medicine.
💻 We introduce scalable and controllable methods for generating natural language instructions robustness evaluation in different settings.
👉Based on these responses, we think there are ~very exciting research opportunities~ in the space of interactive explanations!👈
We propose a set of 5 principles such systems should follow...
@ben_golub
This is a graph I showed admitted grad students in the last visit day. The point was that they will never have as much confidence as they do right now, but with time they will regain ~75% of it back.
How can you guarantee the correctness of each individual prediction? New work with
@StefanoErmon
(AISTATS'21 oral) provides a new perspective on this age-old dilemma based on ideas like insurance and game theory.
Blog:
Arxiv:
Considering these principles, we suggest *natural language dialogues* as a highly promising way to achieve interactive explanations!
🤔What could such a system look like?
❓How could you achieve it?