مقاله انگلیسی رایگان در مورد طبقه بندی تصاویر به عنوان آمار توصیفی – الزویر 2018

Classification images have become popular tools in psychophysics, yet difficulties associated with their interpretation have often hindered their application. Alternative methods for characterizing perceptual filters have been proposed, and the discussion has often focussed on the degree to which classification images are optimal statistical estimators of system components (e.g. kernels). This technical note argues that those difficulties become irrelevant once the tool is situated within a data-driven interpretational framework. Within this framework, classification images and their nonlinear derivatives are understood not as transparent estimates of system components, but instead as transparent descriptors of data structure. The many pitfalls associated with the former approach, and the power of the latter, are demonstrated via combination of counter-intuitive computer simulations with empirical examples from published literature. A change in perspective over the manner in which this tool is understood and utilized may lead to a more productive engagement with this methodology.

Introduction

What is a classification image, and why is it useful?

We constantly ‘filter’ the world around us. Our sensors (eyes, ears, nose, tongue, skin) are bombarded by signals of various kinds, and our ability to discriminate between two such signals (e.g. blue versus red colours) relies on perceptual filters that retain one signal and throw out the other. Our brain then exploits the activity of many such filters to perform specific actions for the purpose of successfully interacting with the environment. In its simplest account, this filtering process can be summarized by a trace (bellshaped curve in Fig. 1A) that records the response of the perceptual filter (plotted on the y axis) to different values of an environmental characteristic, such as the position of an object along the horizon (plotted on the x axis). From Fig. 1A we infer that this specific filter is selective for objects sufficiently close to the middle position long the x axis (stimulus in Fig. 1C), but stops responding when the object is moved further away from the midpoint (Fig. 1D). The filtering stage outlined above returns a continuous value. Our behavioural decisions, however, do not come in this format: we either decide to run away from a predator, or stay put; we either eat a potentially poisonous food item, or we drop it. In other words, most decisions we take on how we use our sensory representation to interact with the world are discrete (typically binary): we either choose to take an action, or we choose not to. How do we go from our perceptual representation, which comes in the form ‘it is 2× more likely that a predator is hiding behind that bush than not’, to the decision ‘run away!’? The simplest model of how this conversion may happen involves a threshold (Green & Swets, 1966): if the ratio between the likelihood of ‘predator’ versus ‘non-predator’ is greater than some value, e.g. 1, we run away; if it is smaller than that value, we stay put. Because we tend to produce this kind of response somewhat erratically, i.e. our estimate of the likelihood is not always identical under the same environmental conditions due to noise in our sensors and our decisional process (Green, 1964; Neri, 2010a), any model of what decision we take must be itself probabilistic: it can only predict that we will run with probability x. To this end, the model in Fig. 1A converts the output from the filter (on the y axis) onto the probability that it will lead to one of two binary choices (e.g. ‘yes’ versus ‘no’). The ‘link’ function is called a static nonlinearity (an example is shown in Fig. 1B). This function is necessary if one is to re-format the output of the filtering stage into the currency of real-world actions