aicrowd/whestbench-smoke-mlp
收藏资源简介:
WhestBench 2026: ARC White-Box Estimation Challenge数据集是一个用于白盒激活估计的基准测试。具体来说,给定一个小型ReLU多层感知机(MLP)的权重和严格的浮点操作(FLOP)预算,目标是预测当网络输入为标准高斯分布时,每个神经元的平均后激活值。该数据集是WhestBench 2026的训练集,包含预构建的MLP及其通过蒙特卡洛方法计算的真实激活统计信息。每个数据行代表一个MLP,包括权重、每层真实均值(基于1,024个独立高斯输入样本计算)、最终层方差标量(作为诊断来源)和每个MLP的种子(用于确保可重复性)。数据集用于挑战赛,参与者需实现预测方法以估计所有层的均值,并在严格的计算预算下优化最终层的均方误差。数据集还包括模式详细信息、生成方法(使用蒙特卡洛模拟,生产版本使用10^9个样本以确保低误差)、数据集摘要(如1个MLP、宽度256、深度8)以及可重复性和引用信息。
WhestBench 2026: ARC White-Box Estimation Challenge Dataset is a benchmark dataset for white-box activation estimation. Specifically, given the weights of a small ReLU-based multi-layer perceptron (MLP) and a strict floating-point operations (FLOP) budget, the task is to predict the average post-activation value of each neuron when the network input follows a standard Gaussian distribution. This dataset is the training set of WhestBench 2026, which contains pre-built MLPs and their ground-truth activation statistics calculated via the Monte Carlo method. Each data row represents an MLP, including its weights, the ground-truth mean of each layer (calculated based on 1,024 independent Gaussian input samples), the final-layer variance scalar for diagnostic purposes, and the random seed of each MLP to ensure reproducibility. This dataset is designed for the challenge, where participants are required to implement prediction methods to estimate the mean values of all layers, and optimize the mean squared error (MSE) of the final layer under the strict computational budget. The dataset also includes detailed information on data patterns, generation method (utilizing Monte Carlo simulation, with 10^9 samples used in the production version to ensure low estimation error), dataset summaries (e.g., an MLP with width 256 and depth 8), as well as reproducibility and citation information.




