A strategy for presenting computational models intelligibly
A strategy for presenting computational models intelligibly
复制标题
一种清晰地呈现计算模型的策略
DOI：
10.1167/jov.21.9.2547
复制
发表时间：
2021
影响因子：
1.8
通讯作者：
Maertens
中科院分区：
文献类型：
--
作者：
Pohlmann;M. Schmittwilken;Maertens
计算模型是一个有用的工具来表征视觉感知的机制。它们避免了语言模型描述中固有的模糊性，并且当与代码一起发布时，它们可以被不同的人（重新）使用，并且它们的预测可以直接复制。出于这个原因，许多期刊现在要求作者将他们的代码与论文一起发布。虽然这是一个值得称赞的做法，但我们认为这还不够，因为几乎没有软件工程背景的读者可能仍然很难将发布的代码与论文中的理论概念联系起来。随着我们对感知过程的理解不断加深，相应的模型也变得越来越复杂。例如，通过时间维度扩展纯空间模型会给这些模型增加显著的复杂性。多维结构的心理和计算处理客观上是困难的，这使得读者很难理解相关的模型部分，更不用说评估它们的适当性了。我们建议使用交互式编程工具（Xueyter）来展示模型。这些工具自然地弥合了文本中的数学模型描述与代码中的相应函数之间的差距。我们选择了一个模型，它结合了视网膜神经节细胞的时空处理特性，但仍然相对简单。它在最近的一篇论文（2019年）中发表在一本明确鼓励发布代码的期刊上。我们单独呈现和可视化模型的功能组件，以及它们各自的输入和输出，帮助读者理解组件及其相互作用。交互式工具的使用允许读者修改参数并观察所产生的效果。尽管尽了最大的努力，我们在原作者的介绍中遇到了一些模糊之处，这些模糊之处可以通过我们在这里提倡的更全面的模型介绍方法来避免。
Computational models are a useful tool to characterize the mechanisms underlying visual perception. They avoid the ambiguity inherent in verbal model descriptions, and, when published together with the code, they can be (re-) used by different people and their predictions can be replicated in a straight-forward way. For this reason, many journals now require authors to publish their code alongside the paper. While this is a commendable practice, we think it is not yet sufficient, because readers with little background in software engineering might still find it difficult to connect the published code with the theoretical concepts in the paper. With our increased understanding of perceptual processes, the corresponding models become more and more complex. For example, extending purely spatial models by a temporal dimension adds significant complexity to such models. The mental and computational handling of multidimensional structures is objectively difficult and makes it hard for readers to understand relevant model parts, let alone assess their adequacy. We suggest to showcase models accessibly using interactive programming tools (Jupyter). These tools naturally bridge the gap between mathematical model descriptions in the text and corresponding functions in the code. We chose a model that incorporates spatio-temporal processing characteristics of retinal ganglion cells but is still relatively straight-forward. It was presented in a recent paper (2019) in a journal that explicitly encourages the publication of code. We present and visualize functional components of the model individually, together with their respective in-and outputs, helping the reader to understand the components and their interactions. The use of an interactive tool allows the reader to tinker with parameters and observe the resulting effects. Despite best efforts, we encountered a number of ambiguities in the original authors' presentation which can be avoided with a more comprehensive approach of model presentation as the one we advocate here.