NeuroEvoBench

NeuroEvoBench是一个专为深度学习应用设计的进化优化算法基准数据集，由柏林工业大学智能科学集群创建。该数据集包含11个不同类别的优化任务，覆盖了广泛的深度学习问题，如视觉分类、控制任务和序列预测等。数据集的创建过程利用了最新的硬件加速和软件框架，如JAX和PyTorch，以支持大规模的并行评估。NeuroEvoBench旨在解决深度学习中复杂的优化问题，特别是那些梯度下降方法难以处理的问题，如非可微分算子的优化和长计算图的梯度计算。

NeuroEvoBench is a benchmark dataset for evolutionary optimization algorithms tailored for deep learning applications, developed by the Cluster of Excellence in Intelligent Sciences at the Technical University of Berlin. This dataset encompasses 11 distinct categories of optimization tasks, covering a wide range of deep learning problems such as visual classification, control tasks, sequence prediction, and more. The development of this dataset leverages state-of-the-art hardware acceleration and software frameworks including JAX and PyTorch to support large-scale parallel evaluation. NeuroEvoBench aims to address complex optimization problems in deep learning, particularly those that are difficult for gradient descent methods to handle, such as optimization of non-differentiable operators and gradient computation for long computational graphs.