Improved Bayes Risk Can Yield Reduced Social Welfare Under Competition

Authors: Meena Jagadeesan, Michael Jordan, Jacob Steinhardt, Nika Haghtalab

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 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.
Researcher Affiliation Academia Meena Jagadeesan UC Berkeley mjagadeesan@berkeley.edu Michael I. Jordan UC Berkeley jordan@cs.berkeley.edu Jacob Steinhardt UC Berkeley jsteinhardt@berkeley.edu Nika Haghtalab UC Berkeley nika@berkeley.edu
Pseudocode No The paper describes algorithms (e.g., best-response dynamics) in text but does not include formal pseudocode blocks or algorithm listings.
Open Source Code Yes The code can be found at https://github.com/mjagadeesan/competition-nonmonotonicity.
Open Datasets Yes We consider a binary image classification task on CIFAR-10 [Krizhevsky, 2009] with 50,000 images... Representations are generated from five models Alex Net [Krizhevsky et al., 2012], VGG16 [Simonyan and Zisserman, 2015], Res Net18, Res Net34, and Res Net50 [He et al., 2016] pretrained on Image Net [Deng et al., 2009].
Dataset Splits Yes We treat the set of 50,000 images and labels as the population of users, meaning that it is both the training set and the validation set.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as CPU or GPU models.
Software Dependencies No The paper does not specify version numbers for any software dependencies or libraries used.
Experiment Setup Yes In more detail, for each j ∈ [m], we initialize the model parameters φ as mean zero Gaussians with standard deviation σ... We implement the approximate best-response as running several steps of gradient descent... We compute the equilibria as follows. First, we let D be the empirical distribution over N = 10, 000 samples from the continuous distribution. Then we run the best-response dynamics described in Section 4.1 with ε = ε0 = 1, I = 5000, δ = 0.1, η = 0.001, and σ = 1.0. We also set the noise parameter c in user decisions (1) to 0.3... For CIFAR-10, for m ∈ {3, 4, 5, 6, 8} model-providers with ε = ε0 = 0.3, I = 2000, η = 0.001, σ = 0.5, and a learning rate schedule that starts at α = 1.0.