skip to main content


Title: Improved Bayes Risk Can Yield Reduced Social Welfare Under Competition
As the scale of machine learning models increases, trends such as scaling laws anticipate consistent downstream improvements in predictive accuracy. However, these trends take the perspective of a single model-provider in isolation, while in reality providers often compete with each other for users. In this work, we demonstrate that competition can fundamentally alter the behavior of these scaling trends, even causing overall predictive accuracy across users to be non-monotonic or decreasing with scale. We define a model of competition for classification tasks, and use data representations as a lens for studying the impact of increases in scale. We find many settings where improving data representation quality (as measured by Bayes risk) decreases the overall predictive accuracy across users (i.e., social welfare) for a marketplace of competing model-providers. Our examples range from closed-form formulas in simple settings to simulations with pretrained representations on CIFAR-10. At a conceptual level, our work suggests that favorable scaling trends for individual model-providers need not translate to downstream improvements in social welfare in marketplaces with multiple model providers.  more » « less
Award ID(s):
2145898
NSF-PAR ID:
10494287
Author(s) / Creator(s):
; ; ;
Publisher / Repository:
Advances in Neural Information Processing Systems 36 (NeurIPS 2023)
Date Published:
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract

    How does an increase in market size, say due to globalization, affect welfare? We study this question using a model with monopolistic competition, heterogeneous markups, and fixed costs. We characterize changes in welfare and decompose changes in allocative efficiency into three different effects: (1) reallocations across firms with heterogeneous price elasticities due to intensifying competition, (2) reallocations due to the exit of marginally profitable firms, and (3) reallocations due to changes in firms’ markups. Whereas the second and third effects have ambiguous implications for welfare, the first effect, which we call the Darwinian effect, always increases welfare regardless of the shape of demand curves. We nonparametrically calibrate demand curves with data from Belgian manufacturing firms and quantify our results. We find that mild increasing returns at the microlevel can catalyze large increasing returns at the macrolevel. Between 70 and 90% of increasing returns to scale come from improvements in how a larger market allocates resources. The lion’s share of these gains are due to the Darwinian effect, which increases the aggregate markup and concentrates sales and employment in high-markup firms. This has implications for policy: an entry subsidy, which harnesses Darwinian reallocations, can improve welfare even when there is more entry than in the first best.

     
    more » « less
  2. Unlicensed spectrum has been viewed as a way to increase competition in wireless access and promote innovation in new technologies and business models. However, several recent papers have shown that the openness of such spectrum can also lead to it becoming over congested when used by competing wireless service providers (SPs). This in turn can result in the SPs making no profit and may deter them from entering the market. However, this prior work assumes that unlicensed access is a separate service from any service offered using licensed spectrum. Here, we instead consider the more common case were service providers bundle both licensed and unlicensed spectrum as a single service and offer this with a single price. We analyze a model for such a market and show that in this case SPs are able to gain higher profit than the case without bundling. It is also possible to get higher social welfare with bundling. Moreover, we explore the case where SPs are allowed to manage the customers' average percentage of time they receive service on unlicensed spectrum and characterize the social welfare gap between the profit maximizing and social welfare maximizing setting. 
    more » « less
  3. Work on scaling laws has found that large language models (LMs) show predictable improvements to overall loss with increased scale (model size, training data, and compute). Here, we present evidence for the claim that LMs may show inverse scaling, or worse task performance with increased scale, e.g., due to flaws in the training objective and data. We present empirical evidence of inverse scaling on 11 datasets collected by running a public contest, the Inverse Scaling Prize, with a substantial prize pool. Through analysis of the datasets, along with other examples found in the literature, we identify four potential causes of inverse scaling: (i) preference to repeat memorized sequences over following in-context instructions, (ii) imitation of undesirable patterns in the training data, (iii) tasks containing an easy distractor task which LMs could focus on, rather than the harder real task, and (iv) correct but misleading few-shot demonstrations of the task. We release the winning datasets at inversescaling.com/data to allow for further investigation of inverse scaling. Our tasks have helped drive the discovery of U-shaped and inverted-U scaling trends, where an initial trend reverses, suggesting that scaling trends are less reliable at predicting the behavior of larger-scale models than previously understood. Overall, our results suggest that there are tasks for which increased model scale alone may not lead to progress, and that more careful thought needs to go into the data and objectives for training language models. 
    more » « less
  4. Work on scaling laws has found that large language models (LMs) show predictable improvements to overall loss with increased scale (model size, training data, and compute). Here, we present evidence for the claim that LMs may show inverse scaling, or worse task performance with increased scale, e.g., due to flaws in the training objective and data. We present empirical evidence of inverse scaling on 11 datasets collected by running a public contest, the Inverse Scaling Prize, with a substantial prize pool. Through analysis of the datasets, along with other examples found in the literature, we identify four potential causes of inverse scaling: (i) preference to repeat memorized sequences over following in-context instructions, (ii) imitation of undesirable patterns in the training data, (iii) tasks containing an easy distractor task which LMs could focus on, rather than the harder real task, and (iv) correct but misleading few-shot demonstrations of the task. We release the winning datasets at https://inversescaling.com/data to allow for further investigation of inverse scaling. Our tasks have helped drive the discovery of U-shaped and inverted-U scaling trends, where an initial trend reverses, suggesting that scaling trends are less reliable at predicting the behavior of larger-scale models than previously understood. Overall, our results suggest that there are tasks for which increased model scale alone may not lead to progress, and that more careful thought needs to go into the data and objectives for training language models. 
    more » « less
  5. Abstract

    We model multistage competition in supply chains by a sequential game theoretical model in a network. We show the uniqueness of equilibrium for series parallel networks and provide a linear time algorithm to compute the equilibrium. We study the impact of the network structure on the firms' payoffs, the total trade flow, and the overall social welfare.

     
    more » « less