微信自助辅助验证机器人_机器学习辅助构想的进步

最新推荐文章于 2024-08-18 15:02:35 发布

weixin_26630173

最新推荐文章于 2024-08-18 15:02:35 发布

阅读量394

点赞数

文章标签：人工智能小程序 python 机器学习 java ViewUI

原文链接：https://medium.com/@adampickard_44261/advancements-in-machine-learning-assisted-ideation-5c42cdf69c37

版权

微信自助辅助验证机器人

The most common debate around Artificial Intelligence and Machine Learning is “Will AI Take Your Job — or Make It Better?.” If most people had a choice, they would probably choose the latter. With any of these new technologies, it can be challenging to distinguish the hype from the headline. On one end, you have big tech companies and startups promising to fix problems ranging from detecting cancer to creating COVID vaccines. On the other, there are people’s understandable fear of becoming obsolete in the workplace.

围绕人工智能和机器学习的最普遍辩论是“ 人工智能会接手您的工作还是使其变得更好？。” 如果大多数人都有选择，他们可能会选择后者。使用这些新技术中的任何一种，要从标题中区分炒作可能是具有挑战性的。一方面，您有大型科技公司和初创公司承诺解决从检测癌症到生产COVID疫苗的问题。另一方面，人们对在工作场所变得过时的恐惧是可以理解的。

If you are unfamiliar with what a GAN is, this video is an excellent short primer.

如果您不熟悉GAN，那么该视频是一个很好的简短入门。

That’s why I read beyond the headlines and use the technology myself when possible. Recently some of these ML technologies have got much more accessible. Apps like Runway give you access to the remote GPU power needed for these tasks along with an ever-growing library of Github repositories while having an interface much closer to something like Photoshop.

这就是为什么我不仅仅关注头条新闻，而且在可能的情况下自己使用这项技术的原因。最近，其中一些机器学习技术变得更加容易使用。诸如Runway之类的应用程序使您可以访问这些任务所需的远程GPU功能，以及不断增长的Github存储库库，同时使界面更接近Photoshop之类的东西。

While taking a class called StyleGAN2 In-Depth, I was excited by a method called Projection that is within the StyleGAN2 library. Something about it got me thinking that ML could be used as a tool for concept development. I’m not claiming that something like this is going to replace human creativity, just to add to the many other existing methods used to help with new ideas.

在学习名为StyleGAN2深入学习的类时，我对StyleGAN2库中称为Projection的方法感到兴奋。关于它的某些事情使我想到ML可以用作概念开发的工具。我并不是说这样的事情会取代人类的创造力，而只是增加了许多其他现有的方法来帮助提出新的想法。

Double-checking, that I’m not a Deepfake! (Projection on the FFHQ model)

仔细检查，我不是Deepfake！ (基于FFHQ模型的投影)

Projection was initially conceived as a method to see if a specific model was used to make a Deepfake image. A possible Deepfake image is selected and the ML model is searched for the closest image it contains in a manner similar to the way google reverse image search can search the web for a specific photo. I’m interested in using it differently. In my mind a creative could train a model with either images of their previous designs or designs by their major influences. My first thought was a creative could use Projection on the resulting model with any design or random image and see if any of the resulting images spark new ideas. The resulting images wouldn’t be the same as the reference image unless included in the original dataset. Adjusting features like truncation settings would cause the results to vary from similar to very wild.

投影最初被认为是一种查看是否使用特定模型制作Deepfake图像的方法。选择了一个可能的Deepfake图像，然后以类似于google反向图像搜索可以在网络上搜索特定照片的方式搜索ML模型以查找其中包含的最接近图像。我有兴趣以其他方式使用它。在我看来，一个创意者可以用先前设计的图像或受到主要影响的设计来训练模型。我的第一个想法是，创意人员可以在具有任何设计或随机图像的结果模型上使用Projection，并查看是否有任何结果图像激发了新的想法。除非原始数据集中包含该图像，否则所得图像将与参考图像不同。调整诸如截断设置之类的功能会导致结果从相似到非常狂野。

The image dataset used to train and latent space of the resulting chair model.

用于训练和生成椅子模型的潜在空间的图像数据集。

Some of the chair images used to train my model and resulting sequence.

一些椅子图像用于训练我的模型和结果序列。

I chose to make my own chair model and use it as an example for two main reasons: in the design world, there is an ongoing debate if there are already too many chairs; and in the ML world, there is a long history of chairs being used for datasets.

我选择制作自己的椅子模型并将其用作示例，主要有两个原因：在设计领域，是否存在太多椅子一直在争论中；在ML世界中，将椅子用于数据集的历史由来已久。

While writing this, I discovered Philipp Schmitt and Steffen Weiß’s The Chair Project. Their project is similar enough to what I was suggesting that I changed this post’s focus to look at some of the recent ML advances since their project was published. While it has been just two years since they wrote their paper, you can see that the realism of the generated images has come along way, as have some of the tools for interacting with a model.

在撰写本文时，我发现了Philipp Schmitt和SteffenWeiß的The Chair Project 。他们的项目足够类似于我所建议的，我改变了这篇文章的重点，以查看自他们的项目发布以来最近在机器学习方面的一些进步。他们撰写论文仅两年时间，但是您可以看到，所生成图像的逼真度以及与模型进行交互的一些工具已经实现。

Also, even though this focuses on chair design, I believe methods like this could be used for a wide range of creative fields. For example, graphic designers could explore patterns, type, textures and motion. Possible for any creative with a good dataset of images.

同样，即使这是针对椅子设计的，我相信这样的方法也可以用于广泛的创意领域。例如，图形设计师可以探索图案，类型，纹理和运动。任何具有良好图像数据集的广告素材都可以使用。

Specific to our chair example, there are products currently available that were made with generative design methods. A good example is Philippe Starck’s A.I. chair for Kartell. Using ML to generate the optimum amount and form for any specific material has obvious advantages for the environment and production efficiencies, but the downside is a form that is strangely both organic and alien at the same time.

特定于我们的椅子示例，当前有一些可用生成设计方法制造的产品。 Philippe Starck担任Kartell的AI主席就是一个很好的例子。使用ML生成任何特定材料的最佳量和形式对于环境和生产效率具有明显的优势，但不利的是同时存在有机物和异物的形式。

Much of the GAN related artwork you see out there involves some latent space animation. It’s an excellent way to gauge a models’ success but doesn’t allow human interaction.

您在那里看到的许多GAN相关作品都涉及一些潜在的空间动画。这是评估模型成功的极好方法，但不允许人工干预。

Here are a couple of those the current methods.

这是目前的几种方法。

Projection method used to compare my chair model against famous designs.

用于将我的椅子模型与著名设计进行比较的投影方法。

An example of Projection in action. The image could an existing model or sketch, and generated on the right is the closest design the model can find.

投影在行动中的一个例子。该图像可以是现有的模型或草图，并且在右侧生成的是该模型可以找到的最接近的设计。

GANSpace GUI made using my model.

使用我的模型制作的GANSpace GUI。

GANSpace allows for the use of a GUI with your model. You are given up to 512 options called Components. These options are how the computer understands various differences, so many don’t make sense to us. You can label them and use them in an interface as above.

GANSpace允许在模型中使用GUI。最多提供512个称为“组件”的选项。这些选项是计算机理解各种差异的方式，因此许多对我们来说都是没有意义的。您可以标记它们并在上面的界面中使用它们。

Example of people yelling at their GAN using Audio-Reactive.

人们使用Audio-Reactive在GAN上大喊大叫的示例。

Audio-reactive allows a user to explore your model through sound. These can be random or chosen vectors in your model. So why not yell or sing to your model?

音频React性使用户可以通过声音探索您的模型。这些可以是模型中的随机向量或选定向量。那么，为什么不大喊或唱歌呢？

Image for post — Chair styles are mixed room the top and left.

Mix Seeds is a method for combining two seeds together. I’ve seen examples of it working better for other people, but it mostly just transferred the colour for me.

混合种子是将两个种子组合在一起的方法。我已经看到了一些示例，它对其他人更有效，但大多数情况下只是为我转移了颜色。

What I hoped Mix Seeds would do it looks like StarGAN v2 would do better. I wasn’t able to use my model with it, but if the sample animation is representative of its capabilities, I could see it being very useful.

我希望Mix Seeds可以做到，就像StarGAN v2会做得更好。我无法将其与模型一起使用，但是如果示例动画可以代表其功能，则可以看到它非常有用。

Me testing out the ergonomics of my model.

我测试了模型的人体工程学。

One thing is clear, a human is still needed to consider materials, manufacturing, and obviously, ergonomics as my video above demonstrates:)

一件事很清楚，正如我上面的视频所示，人类仍然需要考虑材料，制造以及显然的人体工程学：)

Some ML Links to get you started:Runway & youtube channel Derrick Schultz’s youtube channel — One of my instructorsOur Class’s StyleGAN2 Show and Tell

一些ML链接可帮助您入门： 跑道和youtube频道 Derrick Schultz的youtube频道 -我的老师之一我们班的StyleGAN2显示和讲述

Some other ML links worth checking out:Deep Learning In 5 Minutes Techniques in Self-Attention Generative Adversarial Networks Two Minute Papers This chair does not exist The Internet Furry Drama Raising Big Questions About Artificial Intelligence When audio deepfakes put words in Jay-Z’s mouth, did he have a legal case?

其他一些ML链接值得检查： 自我注意生成对抗网络中的 5分钟深度学习技术两分钟论文不存在这把椅子互联网毛茸茸的戏剧引发了关于人工智能的大问题当音频假冒东西在Jay-Z的嘴里放话时，他有法律诉讼吗？