Upcoming
29 October 2024 (Tuesday)
- Where/When: JBHT 535, 12:30-1:30P (U of A)
-
Reading:
- Selective Generation for Language Models - Minjae Lee, Kyungmin Kim, Taesoo Kim, Sangdon Park
- Meeting Link (opt): email me!
Abstract: Trustworthiness of generative language models (GLMs) is crucial due to the recent surge of GLMs in critical decision making systems. Conventional certified uncertainty learning methods, including selective classification and conformal prediction, have been gradually addressing the trustworthy issues of classical models but they still lack for mitigating the metric misalignment issue in GLMs, i.e., good metrics for language generation tasks in measuring the semantic correctness of the true and generated answers are missing. In this paper, we leverage the concept of logical entailment to identify the relation between the true and generated answers to define an entailment-based false discovery rate (FDR) and propose supervised selective generation that exploits entailment labels to control the FDR. As obtaining the entailment labels is expensive, we mitigate this by proposing semi-supervised selective generation that leverages samples without entailment labels by learning an entailment set, which plays a pseudo labeling function, and by designing neuro-selection functions, which further enhances a data space for entailment set learning to minimize the FDR. The proposed neuro-selective entailing-generation algorithm (NSeGen) is theoretically justified to provide the correctness guarantee in the entailment-based FDR, where we also provide sufficient conditions to always satisfy the PAC guarantee on the desired FDR. We demonstrate the efficacy of NSeGen in achieving a desired level of the FDR and better efficiency compared to baselines on open and closed source GLMs.
Past Reading Groups
-
AI Art is Theft: Labour, Extraction, and Exploitation Or, On the Dangers of Stochastic Pollocks
Check out the paper AI Art is Theft: Labour, Extraction, and Exploitation Or, On the Dangers of Stochastic Pollocks Trystan S. Goetze Abstract Since the launch of applications such as DALL•E, Midjourney, and Stable Diffusion, generative artificial intelligence has been controversial as a tool for creating artwork. Some writers have presente... >
-
Towards A Rigorous Science of Interpretable Machine Learning
Check out the paper Towards A Rigorous Science of Interpretable Machine Learning Finale Doshi-Velez and Been Kim Abstract As machine learning systems become ubiquitous, there has been a surge of interest in interpretable machine learning: systems that provide explanation for their outputs. These explanations are often used to qualitatively ... >
-
Machine Unlearning
Check out the paper Machine Unlearning Abstract Once users have shared their data online, it is generally difficult for them to revoke access and ask for the data to be deleted. Machine learning (ML) exacerbates this problem because any model trained with said data may have memorized it, putting users at risk of a successful privacy attack ... >
-
AI Art and its Impact on Artists
Check out the paper AI Art and its Impact on Artists Abstract The last 3 years have resulted in machine learning (ML)-based image generators with the ability to output consistently higher quality images based on natural language prompts as inputs. As a result, many popular commercial “generative AI Art” products have entered the market, mak... >
-
Evaluating the Impact of Local Differential Privacy on Utility Loss via Influence Functions
Check out the paper Evaluating the Impact of Local Differential Privacy on Utility Loss via Influence Functions Abstract How to properly set the privacy parameter in differential privacy (DP) has been an open question in DP research since it was first proposed in 2006. In this work, we demonstrate the ability of influence functions to offer... >
-
Probabilistic Dataset Reconstruction from Interpretable Models
Check out the paper Probabilistic Dataset Reconstruction from Interpretable Models Check out the slides PDF here. >
-
Risks From Learned Optimization in Advanced Machine Learning Systems
Check out the paper Risks From Learned Optimization in Advanced Machine Learning Systems Check out the slides >
-
Causal Fairness Field Guide: Perspectives from Social and Formal Sciences
Check out the paper Causal Fairness Field Guide: Perspectives from Social and Formal Sciences Causality Chapter from the Fair ML Book >
-
Underspecification Presents Challenges for Credibility in Modern Machine Learning
Check out the paper MIT Technology Review Article Underspecification Presents Challenges for Credibility in Modern Machine Learning Check out the slides PDF here. >
-
Arbitrariness and Social Prediction: The Confounding Role of Variance in Fair Classification
Check out the paper Arbitrariness and Social Prediction: The Confounding Role of Variance in Fair Classification - A. Feder Cooper, K. Lee, M. Zahrah Choksi, S. Barocas, C. De Sa, J. Grimmelmann, J. Kleinberg, S. Sen, B. Zhang Check out the slides PDF here. >
-
Runaround + Code 8
Check out the Story and Short Runaround - Isaac Asimov Code 8 (Youtube) Short Film note: Code 8 was turned into a full length feature (now on Netflix). >
-
Runaround - Isaac Asimov
Check out the paper Runaround by Isaac Asimov Check out the slides PDF here. >
-
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
Check out the paper Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training - Evan Hubinger, Carson Denison, Jesse Mu, Mike Lambert, Meg Tong, Monte MacDiarmid, Tamera Lanham, Daniel M. Ziegler, Tim Maxwell, Newton Cheng, Adam Jermyn, Amanda Askell, Ansh Radhakrishnan, Cem Anil, David Duvenaud, Deep Ganguli, Fazl Barez, J... >
-
Model Explanations with Differential Privacy
Check out the paper Model Explanations with Differential Privacy - N. Patel, R. Shokri, and Y. Zick. 2022 Check out the slides PDF here. >
-
Causal parrots: Large language models may talk causality but are not causal
Check out the paper Causal parrots: Large language models may talk causality but are not causal - Zečević, M., Willig, M., Dhami, D. S., & Kersting, K. (2023) Check out the slides PDF here. >
-
Differentially Private Fair Learning
Check out the paper Differentially Private Fair Learning - Matthew Jagielski, Michael Kearns, Jieming Mao, Aaron Oprea, Alina Roth, Saeed Sharifi-Malvajerdi, and Jonathan Ullman Check out the slides PDF here. >
-
Welcome / Intro
This week we covered basic expectations and topics we hope to cover. Nothing big here! >