Who Validates the Validators? Aligning LLM-Assisted Evaluation of LLM Outputs with Human Preferences

Due to the cumbersome nature of human evaluation and limitations of code-based evaluation, Large Language Models (LLMs) are increasingly being used to assist humans in evaluating LLM outputs. Yet LLM-generated evaluators simply inherit all the problems of the LLMs they evaluate, requiring further human validation. We present a mixed-initiative approach to ``validate the validators'' -- aligning LLM-generated evaluation functions (be it prompts or code) with human requirements. Our interface, EvalGen, provides automated assistance to users in generating evaluation criteria and implementing assertions. While generating candidate implementations (Python functions, LLM grader prompts), EvalGen asks humans to grade a subset of LLM outputs; this feedback is used to select implementations that better align with user grades. A qualitative study finds overall support for EvalGen but underscores the subjectivity and iterative process of alignment. In particular, we identify a phenomenon we dub \emph{criteria drift}: users need criteria to grade outputs, but grading outputs helps users define criteria. What is more, some criteria appears \emph{dependent} on the specific LLM outputs observed (rather than independent criteria that can be defined \emph{a priori}), raising serious questions for approaches that assume the independence of evaluation from observation of model outputs. We present our interface and implementation details, a comparison of our algorithm with a baseline approach, and implications for the design of future LLM evaluation assistants.

Revisiting Prompt Engineering via Declarative Crowdsourcing

Large language models (LLMs) are incredibly powerful at comprehending and generating data in the form of text, but are brittle and error-prone. There has been an advent of toolkits and recipes centered around so-called prompt engineering-the process of asking an LLM to do something via a series of prompts. However, for LLM-powered data processing workflows, in particular, optimizing for quality, while keeping cost bounded, is a tedious, manual process. We put forth a vision for declarative prompt engineering. We view LLMs like crowd workers and leverage ideas from the declarative crowdsourcing literature-including leveraging multiple prompting strategies, ensuring internal consistency, and exploring hybrid-LLM-non-LLM approaches-to make prompt engineering a more principled process. Preliminary case studies on sorting, entity resolution, and imputation demonstrate the promise of our approach

Operationalizing Machine Learning: An Interview Study

Organizations rely on machine learning engineers (MLEs) to operationalize ML, i.e., deploy and maintain ML pipelines in production. The process of operationalizing ML, or MLOps, consists of a continual loop of (i) data collection and labeling, (ii) experimentation to improve ML performance, (iii) evaluation throughout a multi-staged deployment process, and (iv) monitoring of performance drops in production. When considered together, these responsibilities seem staggering -- how does anyone do MLOps, what are the unaddressed challenges, and what are the implications for tool builders? We conducted semi-structured ethnographic interviews with 18 MLEs working across many applications, including chatbots, autonomous vehicles, and finance. Our interviews expose three variables that govern success for a production ML deployment: Velocity, Validation, and Versioning. We summarize common practices for successful ML experimentation, deployment, and sustaining production performance. Finally, we discuss interviewees' pain points and anti-patterns, with implications for tool design.

Rethinking Streaming Machine Learning Evaluation

While most work on evaluating machine learning (ML) models focuses on computing accuracy on batches of data, tracking accuracy alone in a streaming setting (i.e., unbounded, timestamp-ordered datasets) fails to appropriately identify when models are performing unexpectedly. In this position paper, we discuss how the nature of streaming ML problems introduces new real-world challenges (e.g., delayed arrival of labels) and recommend additional metrics to assess streaming ML performance.

Enabling certification of verification-agnostic networks via memory-efficient semidefinite programming

Convex relaxations have emerged as a promising approach for verifying desirable properties of neural networks like robustness to adversarial perturbations. Widely used Linear Programming (LP) relaxations only work well when networks are trained to facilitate verification. This precludes applications that involve verification-agnostic networks, i.e., networks not specially trained for verification. On the other hand, semidefinite programming (SDP) relaxations have successfully be applied to verification-agnostic networks, but do not currently scale beyond small networks due to poor time and space asymptotics. In this work, we propose a first-order dual SDP algorithm that (1) requires memory only linear in the total number of network activations, (2) only requires a fixed number of forward/backward passes through the network per iteration. By exploiting iterative eigenvector methods, we express all solver operations in terms of forward and backward passes through the network, enabling efficient use of hardware like GPUs/TPUs. For two verification-agnostic networks on MNIST and CIFAR-10, we significantly improve L-inf verified robust accuracy from 1% to 88% and 6% to 40% respectively. We also demonstrate tight verification of a quadratic stability specification for the decoder of a variational autoencoder.

Optimal Transfer Learning Model for Binary Classification of Funduscopic Images through Simple Heuristics

Deep learning models have the capacity to fundamentally revolutionize medical imaging analysis, and they have particularly interesting applications in computer-aided diagnosis. We attempt to use deep learning neural networks to diagnose funduscopic images, visual representations of the interior of the eye. Recently, a few robust deep learning approaches have performed binary classification to infer the presence of a specific ocular disease, such as glaucoma or diabetic retinopathy. In an effort to broaden the applications of computer-aided ocular disease diagnosis, we propose a unifying model for disease classification: low-cost inference of a fundus image to determine whether it is healthy or diseased. To achieve this, we use transfer learning techniques, which retain the more overarching capabilities of a pre-trained base architecture but can adapt to another dataset. For comparisons, we then develop a custom heuristic equation and evaluation metric ranking system to determine the optimal base architecture and hyperparameters. The Xception base architecture, Adam optimizer, and mean squared error loss function perform best, achieving 90% accuracy, 94% sensitivity, and 86% specificity. For additional ease of use, we contain the model in a web interface whose file chooser can access the local filesystem, allowing for use on any internet-connected device: mobile, PC, or otherwise.

Adversarial Examples that Fool both Computer Vision and Time-Limited Humans

Machine learning models are vulnerable to adversarial examples: small changes to images can cause computer vision models to make mistakes such as identifying a school bus as an ostrich. However, it is still an open question whether humans are prone to similar mistakes. Here, we address this question by leveraging recent techniques that transfer adversarial examples from computer vision models with known parameters and architecture to other models with unknown parameters and architecture, and by matching the initial processing of the human visual system. We find that adversarial examples that strongly transfer across computer vision models influence the classifications made by time-limited human observers.

No Classification without Representation: Assessing Geodiversity Issues in Open Data Sets for the Developing World

Modern machine learning systems such as image classifiers rely heavily on large scale data sets for training. Such data sets are costly to create, thus in practice a small number of freely available, open source data sets are widely used. We suggest that examining the geo-diversity of open data sets is critical before adopting a data set for use cases in the developing world. We analyze two large, publicly available image data sets to assess geo-diversity and find that these data sets appear to exhibit an observable amerocentric and eurocentric representation bias. Further, we analyze classifiers trained on these data sets to assess the impact of these training distributions and find strong differences in the relative performance on images from different locales. These results emphasize the need to ensure geo-representation when constructing data sets for use in the developing world.