Category: Interactive Machine Learning

Posted on

Week 02 Case Study Research – The chAIr Project – Jenny

The chAIr project is a series of four chairs co-designed by human designer and artificial intelligence program. The two artists, Philipp Schmitt and Steffen Weiss want to replicate the classic furniture style of 20th century. They first start with training a generative neural network (GAN) model by incorporating 562 chair designs from Pinterest. Then they set their AI model to only focus on the aesthetic part of each chair, without taking any functional requirements into consideration.

chairs training

After finishing training the model, they run this model and generate sketches as follows. Although these sketches seem to be abstract and may not be used as design drawings directly, the neural network has already implemented common feature of the database into these new sketches. By picking up old classic icons, AI helps us to generate new classic designs.

chairs generating

Then it comes to the part human designers can get involved in. Based on the sketches generated by the machine, human designers can use these design sketches as a starting point to generate their own design sketches. At the same time, human designers can use their imagination and creativity to redesign the blurry parts of AI designed sketches and fix the unreasonable design concepts. 

chair sketches 1

In the end, human designers can finalize their design and start to make prototypes. In this case, Philipp Schmitt and Steffen Weiss 3D-printed 4 chairs generated from their program. We can see the whole design process here starts from the machine and then pass the basic design framework to human designers. Different from what designers do today, this project reverses the role or the sequence human designers and machine play and act during a whole design process. 

chair process 1chair process 2chair process 3chair process 4

The collaborative design process simplifies XX the by training AI to help us stimulate imagination. This kind of collaboration will save time for all kinds of designers and sometimes it may even brings out some surprising outcomes. Philipp Schmitt and Steffen Weiss think that human and AI co-creativity can possibly extend our imaginative capacities in the near future. 

chair design

I really love this project since it can be easily applied into our daily lives and there are many industries that it can be applied into. I am looking forward to seeing this kind of collaborative design happening in the future. 

Designer: Philipp Schmitt & Steffen Weiss

Open Source: DCGAN, arXiv:1511.06434 [cs.LG], Torch Implementation

Project: The Chair Project – Philipp Schmitt

                  The chAIr project – generating a design classic – Steffen Weiss

Posted on

iML Week 2 Case Study: Baidu Translation AI

I ended up reading a whole lot about Machine Learning and AI applications in various fields – gaming, video/photo analysis, face recognition, gaming, especially gaming, agents that game better than humans, etc, but – I chose this option because of its relevance to me and how insanely useful it can be. 

We’ve all used Google translate at some point – where we supply an input and the “translator” generates an output. Most common use for me is when I copy past something in Chinese and hit translate, and I get the output in English. Then, there is the voice function – which I sometimes use to check (note, I said *check*) my Chinese oral homework. In there, you press the little mic button, the “translator” listens to your input and when you press the mic button again, it stops listening and generates an output.

Then there’s this –

The translation is in real time. Baidu claims to use Simultaneous Translation and Anticipation and Controllable Latency (STACL) or whatever that means. The result is that the “translator” is capable of staring translation second after a speaker has begun to speak and end translation momentarily after. To some extent, Baidu’s “translator” is able to anticipate the next word based on the current context and what is being said. In an example that I read about here, a translated word resulted in President Bush in Moscow (in English) and the AI was able to predict that the verb upcoming would be “meet” before the verb was even said (since it’s very likely that Bush will be meeting someone if he is in Moscow?).

There is also a latency feature which can be adjusted based on how closely the languages are related, but I don’t completely get how that works. There’s naturally some quality that’s lost, but it’s still able to produce simultaneous translation, which I think is pretty cool, especially when you read about how it works.  

I don’t have concrete sources like GitHub repos. I’m just going to link the articles I read. Also, here’s a research paper that I didn’t go through entirely(it looked very complex). 

Baidu’s official research blog

Article 1, Article 2

Posted on

IML | Week 2: Case Study Research – Style Transfer for 3D

AI-driven style transfer has been around for many years, the main application of which is to apply a certain style from one image to another. The performance cost has been reduced that even a modern smartphone can run it. And so does high-quality 3D graphics. Is there any similar application for 3D use cases derived from the same technology? The answer is yes, but not quite a lot.

Continue reading “IML | Week 2: Case Study Research – Style Transfer for 3D”

Posted on

Week 2 Case Study Research – BeautyGAN – Zeyao

BeautyGAN is the machine learning model that I mentioned in the first class. It was a perfect example of applying Machine Learning technique to something artistic. Published by 6 researchers, BeautyGAN is an intense-level facial makeup transfer with deep learning generative adversarial network. It transfers make-up style from the makeup dataset to someone who does not wear any makeups. From the paper, it says they “achieve automatic makeup transfer with a dual input/output generative adversarial network” and “achieve instance-level style transfer by successfully applying pixel-level histogram losses on local regions”. It also builds a makeup dataset with 3834 images. 

You can find the paper here

However, what I want to actually talk about is the use of BeautyGAN. I first discovered this machine learning model on Dazed’s Instagram account. Dazed is a British fashion magazine that has a long history. This year, they released their Dazed Beauty Issue 0, the first issue that focuses on the beauty. For the first issue, they got Kylie Jenner, one of the most influential beauty celebrity in the world, to be on the cover. Different from the traditional cover shooting, Dazed Beauty let Beauty AI does Kylie Jenner’s makeup, which brought the BeautyGAN to the public. 

The article explains how would BeautyGAN work in a literal way. “The AI starts out with a data set: 17,000 images pulled from Instagram by the Beauty_GAN team. Those responsible gathered the most
popular and relevant beauty looks they could find, imagery as diverse and colourful as possible, with specs like ‘full face in shot’. They then sorted the imagery into categories and fed it into what is called a discriminator network, where the algorithm begins to learn stereotypical things about the images. It learns to distinguish an eye with make-up from an eye without make-up, or a smiling face versus a frowning face, for example. Eventually, the computer gets so good at distinguishing between categories that it is able to assign categories itself, to differentiate between a beauty selfie or, say, a picture of a dog.”

BeautyGAN provided a new aesthetic for the magazine and the beauty industry. What machine thinks of the beauty and the makeup is uncanny and weird. AI applied trendy makeup to Kylie Jenner, who started the trend, then it had an interesting look on it. When Kylie Jenner started a new trend, millions of people started to copy it. BeautyGAN took all the data and trained the model. When it got applied, it became different than what we thought. Apparently, the machine does not have the same idea as human beings, yet very interesting. 

Posted on

IML: Week 2 Case Study – Thomas Tai

Face classification and detection
Date: February 23rd, 2018

In the past, machines were unable to understand emotions directly, but a breakthrough in the last few years have allowed computers to recognize emotions using machine learning. One of the biggest challenges in designing personalized interaction is understanding the user’s thoughts. For example, a robot assistant would need to monitor a human’s emotions to succeed as a suitable replacement. In the future, computers or phones may use face emotion machine learning to collect data and select suitable content for the user to view. 

One team created a face detection and emotion classification program using Python. According to the report, it can classify gender to a 96% accuracy, and classify emotions at a 66% accuracy. It was trained on a dataset of 460,723 images which defined each person as a woman or man,  and classified each person’s emotions as: “angry”, “disgust”, “fear”, “happy”, “sad”, “surprise”, or “neutral”.

This program is implemented using a convolutional neural networking (CNN) which runs in real time. These are commonly used for image recognition and have been proven to be very effective.  The image is put through a convolutional layer, which is essentially a filter to create a feature map to find matches. The classification uses multiple layers and lots of operations that require training to work. In summary, features are extracted from the image and compared to filters that are determined during training.


Source: medium.com

This project can potentially be very useful in the future as a meaningful form of interaction. Image recognition has revolutionized the way we interact with the world. Google photos uses CNNs to sort pictures by place, type, and person. Every time I pass through Chinese customs or unlock my phone, my face is compared with previous pictures for security purposes. I expect to see many useful purposes for CNNs, and improvements on this algorithm in the future. 

Github Repo: https://github.com/oarriaga/face_classification
Paper: https://github.com/oarriaga/face_classification/blob/master/report.pdf