Why Has Predicting Downstream Capabilities of Frontier AI Models with Scale Remained Elusive?

• URL: http://arxiv.org/abs/2406.04391v2
• Date: Wed, 05 Feb 2025 17:44:38 GMT
• Title: Why Has Predicting Downstream Capabilities of Frontier AI Models with Scale Remained Elusive?
• Authors: Rylan Schaeffer, Hailey Schoelkopf, Brando Miranda, Gabriel Mukobi, Varun Madan, Adam Ibrahim, Herbie Bradley, Stella Biderman, Sanmi Koyejo,
• Abstract summary: This paper identifies a significant factor that makes predicting scaling behavior on question answering benchmarks challenging. We show that downstream performance is computed from negative log likelihoods via a sequence of transformations that progressively degrades the statistical relationship between performance and scale. We study how probability mass on the correct choice co-varies with probability mass on incorrect choices with increasing compute, suggesting that scaling laws for textitincorrect choices might be achievable.
• Score: 26.04581530766348
• License:
• Abstract: Predicting changes from scaling advanced AI systems is a desirable property for engineers, economists, governments and industry alike, and, while a well-established literature exists on how pretraining performance scales, predictable scaling behavior on downstream capabilities remains elusive. While many factors are certainly responsible, this paper identifies a significant factor that makes predicting scaling behavior on widely used multiple-choice question answering benchmarks challenging and illuminates a path towards making such downstream evaluations predictable with scale. Using five model families and twelve well-established multiple-choice benchmarks, we demonstrate that downstream performance is computed from negative log likelihoods via a sequence of transformations that progressively degrades the statistical relationship between performance and scale. We then pinpoint the mechanism causing this degradation: downstream metrics require comparing the correct choice against a small number of specific incorrect choices, meaning accurately predicting downstream capabilities requires predicting not just how probability mass concentrates on the correct choice with scale, but also how probability mass fluctuates on the alternative incorrect choices with scale. We empirically study how probability mass on the correct choice co-varies with probability mass on incorrect choices with increasing compute, suggesting that scaling laws for \textit{incorrect} choices might be achievable. Our work also explains why pretraining scaling laws are commonly regarded as more predictable than downstream capabilities and contributes towards establishing scaling-predictable evaluations of frontier AI models.

Related papers

• Awareness of uncertainty in classification using a multivariate model and multi-views [1.3048920509133808]
  The proposed model regularizes uncertain predictions, and trains to calculate both the predictions and their uncertainty estimations. Given the multi-view predictions together with their uncertainties and confidences, we proposed several methods to calculate final predictions. The proposed methodology was tested using CIFAR-10 dataset with clean and noisy labels.
  arXiv  Detail & Related papers  (2024-04-16T06:40:51Z)
• Selecting Large Language Model to Fine-tune via Rectified Scaling Law [74.84096546112215]
  Given constrained resources, fine-tuning all models and making selections afterward is unrealistic. We find that the fine-tuning scaling curve includes not just the well-known "power phase" but also the previously unobserved "pre-power phase" By leveraging our law, we propose a novel LLM selection algorithm that selects the near-optimal model with hundreds of times less resource consumption.
  arXiv  Detail & Related papers  (2024-02-04T01:55:00Z)
• Improved Bayes Risk Can Yield Reduced Social Welfare Under Competition [99.7047087527422]
  In this work, we demonstrate that competition can fundamentally alter the behavior of machine learning scaling trends. We find many settings where improving data representation quality decreases the overall predictive accuracy across users. At a conceptual level, our work suggests that favorable scaling trends for individual model-providers need not translate to downstream improvements in social welfare.
  arXiv  Detail & Related papers  (2023-06-26T13:06:34Z)
• Calibrated Selective Classification [34.08454890436067]
  We develop a new approach to selective classification in which we propose a method for rejecting examples with "uncertain" uncertainties. We present a framework for learning selectively calibrated models, where a separate selector network is trained to improve the selective calibration error of a given base model. We demonstrate the empirical effectiveness of our approach on multiple image classification and lung cancer risk assessment tasks.
  arXiv  Detail & Related papers  (2022-08-25T13:31:09Z)
• Uncertainty estimation of pedestrian future trajectory using Bayesian approximation [137.00426219455116]
  Under dynamic traffic scenarios, planning based on deterministic predictions is not trustworthy. The authors propose to quantify uncertainty during forecasting using approximation which deterministic approaches fail to capture. The effect of dropout weights and long-term prediction on future state uncertainty has been studied.
  arXiv  Detail & Related papers  (2022-05-04T04:23:38Z)
• Taming Overconfident Prediction on Unlabeled Data from Hindsight [50.9088560433925]
  Minimizing prediction uncertainty on unlabeled data is a key factor to achieve good performance in semi-supervised learning. This paper proposes a dual mechanism, named ADaptive Sharpening (ADS), which first applies a soft-threshold to adaptively mask out determinate and negligible predictions. ADS significantly improves the state-of-the-art SSL methods by making it a plug-in.
  arXiv  Detail & Related papers  (2021-12-15T15:17:02Z)
• Calibrating Predictions to Decisions: A Novel Approach to Multi-Class Calibration [118.26862029820447]
  We introduce a new notion -- emphdecision calibration -- that requires the predicted distribution and true distribution to be indistinguishable'' to a set of downstream decision-makers. Decision calibration improves decision-making on skin lesions and ImageNet classification with modern neural network.
  arXiv  Detail & Related papers  (2021-07-12T20:17:28Z)
• Backward-Compatible Prediction Updates: A Probabilistic Approach [12.049279991559091]
  We formalize the Prediction Update Problem and present an efficient probabilistic approach as answer to the above questions. In extensive experiments on standard classification benchmark data sets, we show that our method outperforms alternative strategies for backward-compatible prediction updates.
  arXiv  Detail & Related papers  (2021-07-02T13:05:31Z)
• Multivariate Probabilistic Regression with Natural Gradient Boosting [63.58097881421937]
  We propose a Natural Gradient Boosting (NGBoost) approach based on nonparametrically modeling the conditional parameters of the multivariate predictive distribution. Our method is robust, works out-of-the-box without extensive tuning, is modular with respect to the assumed target distribution, and performs competitively in comparison to existing approaches.
  arXiv  Detail & Related papers  (2021-06-07T17:44:49Z)
• Learnable and Instance-Robust Predictions for Online Matching, Flows and Load Balancing [12.961453245099044]
  We propose a new model for augmenting algorithms with predictions by requiring that they are formally learnable and instance robust. We design online algorithms with predictions for a network flow allocation problem and restricted assignment makespan minimization.
  arXiv  Detail & Related papers  (2020-11-23T21:38:57Z)
• PrognoseNet: A Generative Probabilistic Framework for Multimodal Position Prediction given Context Information [2.5302126831371226]
  We propose an approach which reformulates the prediction problem as a classification task, allowing for powerful tools. A smart choice of the latent variable allows for the reformulation of the log-likelihood function as a combination of a classification problem and a much simplified regression problem. The proposed approach can easily incorporate context information and does not require any preprocessing of the data.
  arXiv  Detail & Related papers  (2020-10-02T06:13:41Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.