认知神经心理学

Understanding and explaining the behavior of deep neural networks remains one of the most important challenges of modern deep learning solutions. The quest for improving the interpretability of deep learning models has moved researchers to draw inspiration from different disciplines such as biology, neuroscience and even psychology. The case for using cognitive psychology to explain neural networks is particularly interesting and it’s the subject of a research paper published by Alphabet’s subsidiary DeepMind last year.

理解和解释深度神经网络的行为仍然是现代深度学习解决方案的最重要挑战之一。对提高深度学习模型的可解释性的追求使研究人员从生物学，神经科学乃至心理学等不同学科中汲取灵感。使用认知心理学来解释神经网络的案例特别有趣，这是Alphabet的子公司DeepMind去年发表的研究论文的主题。

Conceptually, cognitive psychology is the study of different processes of the mind such as attention, language use, memory, perception, problem solving, creativity, and thinking. Some of the original ideas behind cognitive psychology can be traced back to the 1600s but its not until the 1950s when American psychologists challenged some of the traditional ideas behind the dominant school of thought at the time [behaviorism] and adopted a different model of the mind based on advancements in neuroscience. The term “cognitive psychology” is attributed to German-American psychologist Ulric Neisser who in the book of the same title refers to the importance of cognition in behavioral processes:

从概念上讲，认知心理学是对心理的不同过程的研究，例如注意力，语言使用，记忆力，知觉，解决问题的能力，创造力和思维能力。认知心理学背后的一些原始思想可以追溯到1600年代，但是直到1950年代美国心理学家对当时占主导地位的思想流派背后的一些传统思想提出质疑(行为主义)并采用了不同的思维模式基于神经科学的进步。术语“认知心理学”归因于德裔美国人心理学家Ulric Neisser，他在同名书籍中提到认知在行为过程中的重要性：

“The term “cognition” refers to all processes by which the sensory input is transformed, reduced, elaborated, stored, recovered, and used. It is concerned with these processes even when they operate in the absence of relevant stimulation, as in images and hallucinations. … Given such a sweeping definition, it is apparent that cognition is involved in everything a human being might possibly do; that every psychological phenomenon is a cognitive phenomenon. “

“术语”是指所有改变，减少，细化，存储，恢复和使用感觉输入的过程。 即使这些过程在没有相关刺激的情况下(如图像和幻觉)运行，它也与这些过程有关。 …有了这样一个笼统的定义，很明显，人类可能会做的每一件事都涉及到认知。 每个心理现象都是一种认知现象。 “

认知心理学与一站式学习场景 (Cognitive Psychology and One-Shot Learning Scenarios)

Over the last few decades, cognitive psychologists have developed many techniques to explain learning behaviors particularly in children. The DeepMind team borrowed some of those ideas to interpret the behavior of deep neural networks in image classification scenarios. Specifically, DeepMind focuses on cognitive model known as “one-shot learning” that explains the remarkable ability of children to guess the meaning of a word after seeing a single example. As a cognitive phenomenon, one-shot learning is incredibly difficult to interpret. Imagine you are trying to learn a new language and we show you the following image. You would immediately associate that image with a dog and discard all other possible options like “animal”, “puppy”, “white thing”. How did you do that?

在过去的几十年中，认知心理学家已经开发出许多技术来解释学习行为，特别是在儿童中。 DeepMind团队借用了其中的一些想法来解释图像分类场景中的深度神经网络的行为。具体地说，DeepMind专注于称为“一次性学习”的认知模型，该模型解释了儿童在看到单个示例后就能猜测单词含义的非凡能力。作为一种认知现象，一次性学习很难解释。想象一下，您正在尝试学习一种新语言，我们向您展示下图。您将立即将该图像与狗相关联，并丢弃所有其他可能的选项，例如“动物”，“小狗”，“白色事物”。你是怎么做到的？

The answer to one-shot learning scenarios is based on the ability of children to find correct inferences by applying inductive biases to discriminate other options. Based on the paper, some of those inductive biases include:

一键式学习方案的答案基于儿童通过运用归纳性偏见来区别其他选择的能力，从而找到正确的推论。根据该论文，一些归纳偏差包括：

whole object bias, by which children assume that a word refers to an entire object and not its components.

整体对象偏见 ，孩子认为单词是指整个对象而不是其组成部分。
taxonomic bias, by which children assume that a word refers to the basic level category an object belongs to.

分类偏见 ，孩子认为单词是指对象所属的基本级别类别。
shape bias, by which children assume the meaning of a noun is based on object shape rather than color or texture.

形状偏差 ，儿童认为名词含义的依据是对象的形状而不是颜色或纹理。

Knowing this, we need to wonder if one-shot learning deep neural networks that are modeled after children’s cognitive processes exhibit the same biases? To test that hypothesis, the DeepMind team used two widely adopted but poorly understood one-shot learning architectures. Inception Networks are a particular type of convolutional neural network(CNN) focused on one-shot image classification. Matching Networks are a state-of-the-art neural network architecture (also created by DeepMind) that has achieved up to 93% accuracy in one-shot learning scenarios on ImageNet.

知道了这一点，我们需要怀疑以儿童的认知过程为模型的一次性学习型深度神经网络是否表现出相同的偏见？为了验证该假设，DeepMind团队使用了两种被广泛采用但了解程度不高的单发式学习架构。初始网络是一种特殊的卷积神经网络(CNN)，专注于单次图像分类。匹配网络是最先进的神经网络体系结构(也由DeepMind创建)，在ImageNet上的一键式学习方案中，其准确率高达93％。

To understand the bias in one-shot learning networks, the DeepMind team created a dataset of images grouped in three main categories: a probe object, a shape-match object (which is similar to the probe in shape but not in color), and a color-match object (which is similar to the probe in color but not in shape).

为了了解单次学习网络中的偏见，DeepMind团队创建了一个图像数据集，该图像数据集分为三个主要类别：探针对象，形状匹配对象(形状类似于探针，但颜色不相似)，以及颜色匹配的对象(颜色与探针相似，但形状不相似)。

Source: https://arxiv.org/abs/1706.08606

After testing both Inception and Matching Networks with the aforementioned dataset, the research showed a remarkable finding. Just like humans, one-shot learning neural networks have a strong object shape rather than color or texture. In other words, deep neural networks have a ‘shape bias’. From the experiment, DeepMind researchers arrived to the following conclusions:

在使用上述数据集测试了Inception网络和Matching Networks之后，这项研究显示出了惊人的发现。就像人类一样，一次性学习神经网络具有强大的对象形状，而不是颜色或纹理。换句话说，深度神经网络具有“形状偏差”。通过实验，DeepMind研究人员得出以下结论：

Shape bias emerges gradually over the course of early training in the neural networks. This is reminiscent of the emergence of shape bias in humans: young children show smaller shape bias than older children, and adults show the largest bias.在神经网络的早期训练过程中，形状偏差逐渐出现。这使人想起了形状偏见的出现：年幼的孩子比年长的孩子表现出较小的形状偏见，而成年人则表现出最大的偏见。
There are different levels of bias in one-shot learning networks depending on the random seed used for initialization and training. This indicates that is better to use a large sample of trained models to draw valid conclusions when experimenting with deep learning systems, just as psychologists have learnt not to make a conclusion based on a single subject.一键式学习网络中存在不同程度的偏差，具体取决于用于初始化和训练的随机种子。这表明，在尝试深度学习系统时，最好使用大量训练有素的模型得出有效的结论，就像心理学家已经学会了不要基于单个主题得出结论一样。
Neural Networks are able to achieve the same one shot learning performance even when the shape bias is very different, demonstrating that different networks can find a variety of equally effective solutions to a complex problem.即使形状偏差非常不同，神经网络也能够获得相同的一杆式学习性能，这表明不同的网络可以找到解决复杂问题的各种同等有效的解决方案。

The experiments conducted by DeepMind showed important properties of bias in one-shot learning scenarios but, more importantly, it highlighted the value of cognitive psychology techniques to improve the interpretability of neural networks. Who knew deep neural networks needed a shrink

认知神经心理学_使用认知心理学解释神经网络相关推荐

深度学习与计算机视觉系列(9)_串一串神经网络之动手实现小例子
深度学习与计算机视觉系列(9)_串一串神经网络之动手实现小例子作者:寒小阳时间:2016年1月. 出处:http://blog.csdn.net/han_xiaoyang/article/de ...
从任务到可视化，如何理解LSTM网络中的神经元 By 机器之心2017年7月03日 14:29 对人类而言，转写是一件相对容易并且可解释的任务，所以它比较适合用来解释神经网络做了哪些事情，以及神经网
从任务到可视化,如何理解LSTM网络中的神经元 By 机器之心2017年7月03日 14:29 对人类而言,转写是一件相对容易并且可解释的任务,所以它比较适合用来解释神经网络做了哪些事情,以及神经网络 ...
r语言解释回归模型的假设_模型假设-解释
r语言解释回归模型的假设 Ever heard of model assumptions? What are they? And why are they important? A model is ...
收藏 | 图像识别的可视化解释神经网络
点击上方"小白学视觉",选择加"星标"或"置顶" 重磅干货,第一时间送达仅作学术分享,不代表本公众号立场,侵权联系删除转载于:新智元使 ...
形象的解释神经网络激活函数的作用是什么
转载自形象的解释神经网络激活函数的作用是什么神经网络中激活函数的作用查阅资料和学习,大家对神经网络中激活函数的作用主要集中下面这个观点: 激活函数是用来加入非线性因素的,解决性模型所不能解决的 ...
04.卷积神经网络_第一周卷积神经网络
04.卷积神经网络_第一周卷积神经网络 1.1 计算机视觉上面这张图是64x64像素的图像,它的数据量是12288,下图是1000x1000像素的图像,它的数据量是3百万.如果用全连接网络去处理这张 ...
神经网络中神经元是什么_是什么使神经网络脆弱
神经网络中神经元是什么 What do the images below have in common? 以下图片有什么共同点? Most readers will quickly catch on ...
Geoffrey Hinton专访：如何解释神经网络的变迁
https://www.toutiao.com/a6694875591390265860/ 大数据文摘出品来源:wired 编译:蔡婕.顾晨波.张秋玥 Geoffrey Hinton是深度学习的创始 ...
lime 深度学习_用LIME解释机器学习预测并建立信任
lime 深度学习 It's needless to say: machine learning is powerful. 不用说:机器学习功能强大. At the most basic level, ...

认知神经心理学_使用认知心理学解释神经网络

认知心理学与一站式学习场景 (Cognitive Psychology and One-Shot Learning Scenarios)

认知神经心理学_使用认知心理学解释神经网络相关推荐

最新文章

热门文章