node-red 可视化
有关深层学习的FAU讲义 (FAU LECTURE NOTES ON DEEP LEARNING)
These are the lecture notes for FAU’s YouTube Lecture “Deep Learning”. This is a full transcript of the lecture video & matching slides. We hope, you enjoy this as much as the videos. Of course, this transcript was created with deep learning techniques largely automatically and only minor manual modifications were performed. Try it yourself! If you spot mistakes, please let us know!
这些是FAU YouTube讲座“ 深度学习 ”的 讲义 。 这是演讲视频和匹配幻灯片的完整记录。 我们希望您喜欢这些视频。 当然，此成绩单是使用深度学习技术自动创建的，并且仅进行了较小的手动修改。 自己尝试！ 如果发现错误，请告诉我们！
导航 (Navigation)
Previous Lecture / Watch this Video / Top Level / Next Lecture
上一个讲座 / 观看此视频 / 顶级 / 下一个讲座
Welcome back to deep learning! Today, we want to look a bit more into visualization techniques and in particular, the gradient-based and optimization-based procedures.
欢迎回到深度学习！ 今天，我们想更多地研究可视化技术，尤其是基于梯度和基于优化的过程。
Okay, so let’s see what I’ve got for you. Let’s talk first about the gradient-based visualizations and here the idea is that we want to figure out which input pixel is most significant to a neuron.
好吧，让我们看看我能为您带来什么。 让我们先谈谈基于梯度的可视化，这里的想法是我们要弄清楚哪个输入像素对神经元最重要。
If we would change it, what would cause a large variation in the actual output of our neural network? What we actually want to compute is the partial derivative of the neuron under consideration, maybe for an output neuron-like for the class “cat”. Then, we want to compute the partial derivative with respect to the input. This is essentially backpropagation through the entire network. Then, we can visualize this gradient as a type of image which we have been doing here for the cat image. You can see that, of course, this is a color gradient. You see that this is a bit of a noisy image but you can see that what is related to class “cat”, here is obviously also located in the area where the cat is in the image.
如果我们要更改它，什么会导致神经网络的实际输出发生较大变化？ 我们实际上要计算的是所考虑的神经元的偏导数，也许对于“猫”类的输出神经元而言。 然后，我们要计算相对于输入的偏导数。 这实质上是整个网络的反向传播。 然后，我们可以将此梯度可视化为我们在此处为猫图像所做的图像类型。 您当然可以看到，这是一个颜色渐变。 您会看到这有点嘈杂，但可以看到与“猫”类有关的东西，这里显然也位于猫在图像中的区域。
We will learn several different approaches to do this. The first one is based on [20]. For backpropagation, we actually need a loss of what we want to backpropagate. We simply take a pseudo loss that is the activation of an arbitrary neuron or layer. Typically, what you want to do is you want to take neurons in the output layer because they can be associated with a class.
我们将学习几种不同的方法来做到这一点。 第一个基于[20]。 对于反向传播，实际上我们需要丢失要反向传播的内容。 我们简单地采取伪损失，即任意神经元或层的激活。 通常，您想要做的是要在输出层中获取神经元，因为它们可以与类关联。
What you can also do is instead of using backpropagation, you can build a nearly equivalent alternative which uses a kind of reverse network. This is the Deconvnet from [26]. So here, the input is the trained network and some image. Then, you choose one activation and set all of the other activations to zero. Next, you build a reverse network and you can see the idea here that this is essentially containing the same as the network but just in reverse sequence with so-called unpooling steps. Now, with these unpooling steps and the reverse computation, you can see that we can also produce a kind of gradient estimate. The nice thing about this one is, there’s no training involved. So, you just have to record the pooling location in the switches and the forward path of the reverse Network. Effectively this is the same as the backward-pass of the network apart from the rectified linear units which we’ll look at in a couple of slides.
您还可以做的是代替使用反向传播，而可以构建使用反向网络的几乎等效的替代方法。 这是[26]中的去convnet。 因此，这里的输入是经过训练的网络和一些图像。 然后，您选择一个激活并将所有其他激活设置为零。 接下来，您建立了一个反向网络，您可以在这里看到这样的想法：该网络本质上包含与网络相同的内容，但是与反向顺序相同，即所谓的分池步骤。 现在，通过这些解池步骤和反向计算，您可以看到我们还可以生成一种梯度估计。 关于这一点的好处是，无需培训。 因此，您只需要记录交换机中的池位置以及反向网络的正向路径即可。 实际上，这与网络的反向传递相同，除了整流线性单元外，我们将在几张幻灯片中进行介绍。
Here, we show the visualizations of the top nine activations, the gradient, and the corresponding patch. So for example, you can reveal with this one that this kind of feature map seems to focus on green patchy areas. You could argue that this is more a kind of background feature that tries to detect grass patches in the image.
在这里，我们显示了前九个激活，梯度和相应补丁的可视化。 因此，例如，您可以通过此图标揭示这种功能图似乎集中在绿色斑驳的区域。 您可能会说这更多是一种背景功能，试图检测图像中的草斑。
So, what else? Well, there’s guided backpropagation. Guided backpropagation is a very similar concept. The idea here is that you want to find positively correlated features. So we are looking for positive gradients because we assume that the features that are positive are the ones that the neuron is interested in. The negative gradients are the ones that the neuron is not interested in.
还有什么？ 好吧，这里有指导性的反向传播。 引导反向传播是一个非常相似的概念。 这里的想法是您想找到正相关的特征。 因此，我们正在寻找正梯度，因为我们假设正特征是神经元感兴趣的特征。负梯度是神经元不感兴趣的特征。
So, the idea is then to set all negative gradients in the backpropagation to zero. We can show you now the different processes of the ReLU during the forward and backward passes with the different kinds of gradient backpropagation techniques. Well, of course, if you have these input activations, then in the forward pass in the ReLU, you would simply cancel out all the negative values and set them to zero. Now, what happens in the backpropagation for the three different alternatives? Let’s look at what the typical backpropagation does and note that we show here the negative entries that came from the sensitivity in yellow. If you now try to backpropagate this, you have to remember which entries in the forward pass were negative and you set those values again to zero. You keep everything that came from the sensitivity of the previous layer in order to do so. Now if you do Deconvnet, you don’t need to remember the switches from the forward-pass, but you set all the entries that are negative in the sensitivity to zero and backpropagate. This way now, the guided backpropagation actually does both. So it remembers the forward-pass and sets all of those elements to zero. It sets all of the elements of the sensitivities to zero. So, it’s essentially a union of backpropagation and Deconvnet in terms of canceling negative values. You can see that the guided backpropagation only keeps very little sensitivity throughout the entire backpropagation process.
因此，想法是将反向传播中的所有负梯度都设置为零。 现在，我们可以使用不同种类的梯度反向传播技术向您展示正向和反向通过过程中ReLU的不同过程。 好吧，当然，如果您具有这些输入激活，那么在ReLU的正向传递中，您只需取消所有负值并将它们设置为零即可。 现在，对于三种不同的替代方案，反向传播会发生什么？ 让我们看一下典型的反向传播的作用，并注意我们在这里以黄色显示了来自灵敏度的负输入。 如果现在尝试反向传播，则必须记住正向传递中的哪些条目为负，然后将这些值再次设置为零。 为此，您必须保留来自上一层灵敏度的所有内容。 现在，如果您使用Deconvnet，则无需记住前向通行的开关，而是将灵敏度为负的所有条目设置为零并反向传播。 现在，通过这种方式，引导的反向传播实际上可以同时执行这两种操作。 因此，它会记住前向通过并将所有这些元素设置为零。 它将所有灵敏度元素设置为零。 因此，就消除负值而言，它实质上是反向传播和Deconvnet的结合。 您会看到，在整个反向传播过程中，引导反向传播仅保持很小的敏感性。
Now let’s look at the comparison of the different gradients. One thing that you can see is that in Deconvnet, we get pretty noisy activations and backpropagation. We can see that we at least focus on the object of interest and the guided backpropagation has a very sparse representation but you can very clearly see even in this gradient image, the most important features like the eyes of the cat and so on. So, this is a very nice way that might help you reveal which neurons focus on what activity in that specific input.
现在让我们看一下不同梯度的比较。 您可以看到的一件事是，在Deconvnet中，我们得到了非常嘈杂的激活和反向传播。 我们可以看到，我们至少专注于感兴趣的对象，并且引导的反向传播具有非常稀疏的表示，但是即使在此渐变图像中，您也可以非常清晰地看到猫眼等最重要的特征。 因此，这是一种非常好的方法，可以帮助您揭示哪些神经元专注于特定输入中的哪些活动。
This then finally leads to saliency maps. Here, you don’t want to investigate what influences two neurons but you want to investigate the impact of pixels on a class “dog”. So now, you take the pseudo loss as an unnormalized task, compute the gradient with respect to the image pixels, and use absolute values. Then, the interesting observation that we make with this is that it kind of produces a saliency map that localizes the dog in the image, even though the network was never trained on localization. So, this is a very interesting approach that can help you to identify where the decisive information is actually located in the image.
然后，这最终导致显着图。 在这里，您不想调查什么会影响两个神经元，而是想调查像素对“狗”类的影响。 因此，现在，您将伪损失作为未归一化的任务，计算相对于图像像素的梯度，并使用绝对值。 然后，我们由此得出的有趣观察结果是，即使网络从未接受过本地化训练，它仍会产生一种显着性地图，以将狗定位在图像中。 因此，这是一种非常有趣的方法，可以帮助您确定决定性信息在图像中的实际位置。
What else can be done? Well, there’s parameter visualization based on optimization. Now, the idea is that we want to go towards different levels. So, if we want to optimize with respect to a neuron, to an activation map, a layer, the actual logits, or the class probability which is essentially the softmax function, we take them as pseudo loss in order to create optimal inputs.
还有什么可以做的？ 好吧，有基于优化的参数可视化。 现在，我们的想法是要朝着不同的层次发展。 因此，如果我们要针对神经元，激活图，层，实际logit或本质上是softmax函数的类概率进行优化，则可以将它们作为伪损失以创建最佳输入。
We’ve already seen something very similar in the first video where we had this example from DeepDream. Inceptionism is essentially doing something very similar. It takes some input and then it alters the input such that different neurons are maximally activated.
在第一部视频中，我们已经看到了与DeepDream中的示例非常相似的内容。 本质上，盗版主义在做非常相似的事情。 它需要一些输入，然后更改输入，以便最大程度地激活不同的神经元。
There you can see that these neurons somehow encode specific parts of animals or things that it likes to recognize. If you now maximize the input with respect to that particular neuron, you can see that then the shapes that it likes start to appear in this image. So, the idea is that you change the input such that the neuron is maximally activated. So, we are essentially not just computing the gradient up to the image, but we are also actively changing the image with respect to that particular or layer, softmax, or output. The original idea for this was, of course, visualization.
在那里，您可以看到这些神经元以某种方式编码了动物或喜欢识别的事物的特定部分。 如果现在针对该特定神经元最大化输入，则可以看到它喜欢的形状开始出现在此图像中。 因此，您的想法是更改输入，以最大程度地激活神经元。 因此，从本质上讲，我们不仅在计算图像的梯度，而且还针对特定的图层或层，softmax或输出积极地更改图像。 当然，最初的想法是可视化。
So, you try to understand the inner workings of the networks by dreaming about when presented with images. You start with the image or even noise as input. Then, you adjust the image towards maximizing activations in a complete layer. For different layers, it highlights different things in the image. So, we can create this kind of Inceptionism. If you activate mostly early layers, you see that the image content is not that much changed but you create those brush and stroke-like appearances in the images.
因此，您尝试通过梦见何时显示图像来了解网络的内部运作方式。 您从图像甚至噪波开始输入。 然后，您调整图像以最大化整个图层中的激活。 对于不同的图层，它突出显示图像中的不同内容。 因此，我们可以创建这种“感应主义”。 如果您主要激活较早的图层，则会看到图像内容没有太大变化，但是会在图像中创建类似笔刷和笔触的外观。
Now you can even go ahead and start this with random input. Then it’s not just optimizing the input with respect to a specific output. You need some additional regularization. We can show this here in this small formula. So, we are now taking some input x which is a random image. We feed it into our network and a specific neuron or output neuron. Then, we maximize the activation and we add a regularization. This regularizer punishes if our x deviates from a specific norm. What is used in this example, it’s simply the L2 norm. Later, we will also see that maybe also other norms may be suitable for this. So, you start with this noise input that we show on the top right. Then, you optimize until you find a maximum activation for that specific neuron or layer. At the same time, you postulate that your input image somehow has to be smooth because otherwise, you be generating these very, very noisy images. They are not so nice for interpretation and of course, the bottom right image shows you some kind of structures that you know can interpret. So, you see these abstract features emerging and then you can use this as a kind of cascade from small to large scales and this produces the so-called inceptionism.
现在，您甚至可以继续进行，并从随机输入开始。 不仅是针对特定输出优化输入。 您需要一些其他的正则化。 我们可以在这个小公式中显示这一点。 因此，我们现在获取一些输入x ，它是随机图像。 我们将其馈入网络和特定的神经元或输出神经元。 然后，我们最大化激活并添加正则化。 如果我们的x偏离特定范数，则此正则化器将进行惩罚。 在此示例中使用的只是L2规范。 稍后，我们还将看到也许其他规范也可能适用于此。 因此，从我们在右上方显示的此噪声输入开始。 然后，进行优化，直到找到该特定神经元或层的最大激活。 同时，您假定输入图像必须某种程度上必须平滑，因为否则，您将生成这些非常非常嘈杂的图像。 它们不太适合解释，当然，右下角的图像向您展示了您知道可以解释的某种结构。 因此，您看到了这些抽象特征的出现，然后可以将其用作从小到大的级联形式，这会产生所谓的“感应主义”。
Here, we can use that, for example, to reveal hidden weaknesses in the neural network classification process. Here, we see different realizations for the class “dumbbell”. You can see, it’s not only the dumbbell that is shown in the image, but it is also recreating the arm that is holding the dumbbell. So, we can see here that correlated things are kind of learned, when they have been presented to the network. So, we kind of can figure out what the memory of that specific class or neuron with respect to the input is. So, again we learned once more good data is really important.
在这里，我们可以使用它来揭示神经网络分类过程中的隐藏弱点。 在这里，我们看到了“哑铃”类的不同实现。 您会看到，不仅图像中显示了哑铃，而且还重新创建了握住哑铃的手臂。 因此，在这里我们可以看到，当关联事物呈现给网络时，它们就是一种学习的东西。 因此，我们可以确定特定类别或神经元相对于输入的记忆是什么。 因此，我们再次了解到，好的数据确实非常重要。
This actually leads us to another step that we could do in order to figure out what’s happening inside of the neural network. These are inversion techniques and here the idea is very similar to what we’ve seen in the inceptionism idea. But now, we actually want to invert from the activation what was the actual input. What you hear quite frequently, for example, as security measures to anonymize data: “Let’s just take the activations of Layer 5 and discard all the previous activations and inputs. We just store the Layer 5 activations because there is no way how I can reconstruct the original image if I only know that Layer 5 activation.”
实际上，这使我们迈出了另一步，可以弄清楚神经网络内部正在发生什么。 这些是反演技术，这里的想法与我们在感应主义想法中看到的非常相似。 但是现在，我们实际上想从激活中反转什么是实际输入。 例如，您经常听到的作为匿名数据安全措施的信息：“让我们仅进行第5层的激活，并丢弃所有先前的激活和输入。 我们只存储第5层激活，因为如果我仅知道第5层激活，就无法重建原始图像。”
Now with inversion, if you know the network, its processes, and the specific activations for a specific layer, then you can try to reconstruct what the actual input was. So again, we have the output of our network in that particular layer. So let’s say f(x) is the output of a layer and we have y hat. Now, y hat is the measured network’s output or the measured layer activation. So, we have the Layer 5 activation and we don’t know what the input x is. So, we are looking for x and we try to minimize this function such that we find with an x to best match that specific activation.
现在通过反转，如果您知道网络，其过程以及特定层的特定激活，则可以尝试重建实际输入。 同样，我们在该特定层中拥有网络的输出。 假设f( x )是一层的输出，而我们有y hat。 现在，Y帽子是测量网络的输出或所测量的层活化。 因此，我们具有第5层激活功能，我们不知道输入x是什么。 因此，我们正在寻找x，并尝试最小化此函数，以便我们找到一个与x最佳匹配的特定激活。
This is a classical inverse problem. You add in order to get a more stable output, an additional regularizer λ times R(x). This regularizer is something that is very important. So, the regularizer stabilizes the inversion and there are very common techniques for regularization that use specific properties of natural images in order to create something that is likely a natural image. So, of course, high-frequency noise would degrade the reconstructions. This is why we are using this additional L2 norm in order to prevent the appearance of noise in the created images. In addition to that, you can also use the so-called total variation. We know that natural images typically have sparse gradients and total variation is a minimization technique that enforces your image to have a very low number of gradients. Gradients are essentially edges and in a typical image, there are only a few edge pixels and many more homogeneous areas. So, TV minimization produces images with few edges and, of course, also a little noise. It specifically also allows high piecewise constant jumps like in real edges. Of course, you could also work with low-pass and other edge-preserving filters. A classic one is the wavelet regularization. So this is simple, it’s effective and, of course, it will also suppress real edges and other high-frequency information.
这是一个经典的逆问题。 为了获得更稳定的输出，需要添加λ乘以R( x )。 这个正则化器非常重要。 因此，正则化器可稳定反演，并且存在非常常见的正则化技术，这些技术使用自然图像的特定属性来创建可能是自然图像的东西。 因此，当然，高频噪声会降低重建效果。 这就是为什么我们使用此附加L2范数以防止在创建的图像中出现噪点的原因。 除此之外，您还可以使用所谓的总变化。 我们知道自然图像通常具有稀疏的渐变，总变化是一种最小化技术，可以使您的图像具有非常小的渐变数量。 渐变本质上是边缘，在典型图像中，只有很少的边缘像素和更多均匀的区域。 因此，电视最小化产生的图像边缘很少，当然也有少量噪声。 特别是它还允许像实际边缘一样高的分段恒定跳跃。 当然，您也可以使用低通和其他边缘保留滤波器。 经典的一种是小波正则化。 因此，这很简单，很有效，当然也可以抑制真实边缘和其他高频信息。
Well, what else can be done? You can also use other regularizers like transform robustness. So, the input should actually be invariant to special transformation. So this is similar to data augmentation and therefore, you can randomly rotate, scale, or jitter x. So, this is also very simple and it’s effective in producing recognizable features. Often the orientation is suppressed even if it was informative. So, we have to be careful about that.
好吧，还有什么可以做的？ 您还可以使用其他正则化器，例如变换稳健性。 因此，输入实际上应该对特殊变换不变。 因此，这类似于数据扩充，因此，您可以随机旋转，缩放或抖动x 。 因此，这也非常简单，并且可以有效地产生可识别的特征。 即使方向是有益的，通常也会抑制方向。 因此，我们必须对此谨慎。
A last kind of regularizer that is very common is that you have learned priors. So, for example, you can use a train network and say “I want to have a specific distribution in layer #4.” Then, I try to generate images that have a very similar characteristic. Here, instead of optimizing with respect to a specific norm that we know that is useful, we assume that the representations that are produced in a specific layer are useful in order to measure the content of the image. Then, you can actually use this as a kind of regularizer to produce images. So of course, you need a trained generative model if you want to use things like this. This produces very nice images, but it may be ambiguous because parts of what you introduce into the results stem from the pre-trained network. So, you have to see this with a bit of caution.
最后一种很常见的正则化器是您已经了解了先验知识。 因此，例如，您可以使用火车网络并说“我希望在第4层中有特定的分布”。 然后，我尝试生成具有非常相似特征的图像。 在这里，我们不是针对已知的有用规范进行优化，而是假设在特定图层中生成的表示形式对于测量图像的内容很有用。 然后，您实际上可以将其用作一种生成图像的正则化器。 因此，当然，如果您想使用这样的东西，则需要训练有素的生成模型。 这会产生非常漂亮的图像，但是可能会导致模棱两可，因为您引入结果的部分内容来自预先训练的网络。 因此，您必须谨慎一点。
So, let’s look at some examples [14] actually generated images by inversion. This is pretty impressive. Again, this is an AlexNet-type of network and here you have the input and then the inversion:
因此，让我们来看一些通过反演实际生成的图像的示例[14]。 这真是令人印象深刻。 再次，这是AlexNet类型的网络，在这里您有输入，然后是反演：
At the conv layer 1, you can see we can almost exactly reproduce the image. After ReLu 1 not much changed. Pooling — no big effect. Then, the second layer, and so on. You can see that up to convolution layer #4, we are very close to the true input. This has undergone several steps of pooling already and still, we are able to reproduce the input very closely to the original input. Very interesting! Then, you see that I really have to go towards — let’s say — Layer 6 or Layer 7, until I reach a point where it becomes impossible or close to impossible to guess what the original input was. So only, from Layer 6/Layer 7, we start deviating significantly from the original input. Still until Layer 5, we can reconstruct quite well what the original input is. So, if anybody tells you that they want to anonymize data by cutting off the first two layers, then you see that with these inversion techniques this is maybe not such a great idea. It’s not unlikely that you will be able to reconstruct the original input only by means of seeing the activations and the network structure.
在转换层1上，您可以看到我们几乎可以准确地再现图像。 在ReLu 1之后，变化不大。 合并-没什么大作用。 然后，第二层，依此类推。 您可以看到直到卷积层4，我们都非常接近真实输入。 现在已经经历了几个合并的步骤，我们能够非常接近原始输入来重现输入。 很有意思！ 然后，您看到我真的必须走向第6层或第7层，直到到达无法或几乎无法猜测原始输入的地步。 因此，仅从第6层/第7层开始，我们才开始明显偏离原始输入。 直到第5层，我们都可以很好地重建原始输入。 因此，如果有人告诉您他们想通过切断前两层来匿名化数据，那么您会发现，使用这些反演技术可能不是一个好主意。 仅通过查看激活和网络结构就可以重构原始输入。
Okay. So next time, we want to talk about a second topic that is somewhat related to visualization. We want to talk about attention and attention mechanisms. You’ve already seen that with the visualization techniques, we can somehow figure out which pixels are related to what kind of classification. Now, we want to spin this a little further and use this to guide the focus of the attention of the network towards specific areas. So, this will also be a very interesting video. Looking forward to seeing you in the next video. Bye-bye!
好的。 因此，下一次，我们想讨论与可视化有些相关的第二个主题。 我们想谈谈注意力和注意力机制。 您已经看到，借助可视化技术，我们可以以某种方式找出哪些像素与哪种分类有关。 现在，我们想进一步旋转它，并使用它来引导网络关注的焦点转向特定区域。 因此，这也是一个非常有趣的视频。 期待在下一个视频中见到您。 再见！
If you liked this post, you can find more essays here, more educational material on Machine Learning here, or have a look at our Deep LearningLecture. I would also appreciate a follow on YouTube, Twitter, Facebook, or LinkedIn in case you want to be informed about more essays, videos, and research in the future. This article is released under the Creative Commons 4.0 Attribution License and can be reprinted and modified if referenced. If you are interested in generating transcripts from video lectures try AutoBlog.
如果你喜欢这篇文章，你可以找到这里更多的文章 ，更多的教育材料，机器学习在这里 ，或看看我们的深入 学习 讲座 。 如果您希望将来了解更多文章，视频和研究信息，也欢迎关注YouTube ， Twitter ， Facebook或LinkedIn 。 本文是根据知识共享4.0署名许可发布的 ，如果引用，可以重新打印和修改。 如果您对从视频讲座中生成成绩单感兴趣，请尝试使用AutoBlog 。
链接 (Links)
Yosinski et al.: Deep Visualization ToolboxOlah et al.: Feature VisualizationAdam Harley: MNIST Demo
Yosinski等：深度可视化工具箱 Olah等：特征可视化 Adam Harley：MNIST演示
翻译自: https://towardsdatascience.com/visualization-attention-part-4-a1cfefce8bd3
node-red 可视化