AI Favorites Vol.004: Waifu Lab is on fire, how is AI created?

Article

0 1

How does AI draw a virtual wife?

The recently popular Waifu Labs project uses a generative adversarial network method to train AI to create a "virtual wife".

This article [1] dissects the steps of AI learning, we can see how a two-dimensional face is generated in a chaos. Generative Adversarial Networks, or GANs for short, are a method of unsupervised learning. You can think of it as a pair of AIs that play against each other in order to learn: one AI learns how to paint; the other AI learns how to distinguish between paintings made by artificial intelligence and those made by Paintings made by human artists.

By isolating certain eigenvectors, cataloging these coordinates can also enable interesting deformations between characters.

0 2

When AI meets

Sony has been exploring the connection between artificial intelligence and cooking since 2018. A few days ago, the AI system developed by Sony has supported the integration and processing of a large amount of food data and delivered it to the chef, helping the chef to flexibly match the ingredients. In addition, Sony is developing robots that can dexterously and quickly operate a variety of food and cooking utensils to assist chefs in cooking and serving. To achieve these goals, Sony partners with organizations across the industry, from chefs and restaurants to universities, technology companies and food suppliers. We can read further related papers and watch concept videos on Sony's official website [2].

0 3

Meta AI makes children's hand-painted "live"

Children will draw all kinds of strange body shapes, sometimes without a complete head, sometimes with arms sticking out from the eyes... Different children will draw different worlds. How does AI overcome challenges and recognize children's drawings? Meta completes the transition from painting to animation through four steps: object detection to identify human figures; use character masks to extract human figures from the scene; prepare for animation through rigging; use 3D motion capture to animate 3D human figures [3.1]. You can also try it online here [3.2].

0 4

AI software-assisted identification of colorectal

ByteDance's medical brand "Xiaohe Health" has developed an AI-assisted diagnosis software for colonoscopy, which can use artificial intelligence technology to assist clinicians in real-time detection and identification of colorectal cancer lesions.

Compared with other cancers, colorectal cancer is relatively preventable and controllable. The key to prevention and control lies in early prevention, early screening, and early treatment. Colonoscopy is one of the most important early screening methods.

0 5

Jeff Dean: Machine Learning 2021 Summary and Research Trends

Jeff Dean, the head of Google AI, published an annual summary of the 40-character long article. In the summary, five major trends in machine learning are highlighted:

• More capable general-purpose machine learning models
• Continuously improve the efficiency of machine learning
• The benefits of machine learning from a personal and societal perspective
• The benefits of machine learning in science, health and sustainability
• Ethically, Machine Learning Equity and Equity Issues

tool

0 1

AI auto-completion code artifact

🌟 Features:

According to the written context, the matching code is automatically generated, which can automate the simple and repetitive code writing part.

👀 Highlights:

• After training with billions of lines of public code, the actual measurement accuracy rate is high, and it is really cool to press the tab key to automatically complete
• It can also be used to assist software testing, import a software testing package, and automatically complete the rest of the test code

👉Address:

https://copilot.github.com/

**
**

0 2

Creative prototype and audio production tool developed by Sony CSL Music Lab

🌟 Creative features:

• DrumNet realizes drum beat rhythm generation based on existing music melodies through unsupervised learning
• BassNet generates bass based on existing musical melody, and can adjust and control the density, articulation and timbre of the generated notes at any time while the music is playing
• Flow Machines can generate a track pattern with chords, basses and full melodies
• When digitizing the actual piano melody, you often encounter some missing notes. At this time, you can "repair" or even "continue writing" or create a new melody from scratch.
• NOTONO is a synth that visualizes sound
• DrumGAN generates a variety of drum sounds

🔧 Advanced tools for audio processing and production mixing/mastering:

• Profile EQ is similar to the "Auto Contrast" function in a graphic editor, and is an adaptive equalizer that visually compares audio
• Resonance EQ is similar to the "adjust image saturation" function, which can smooth audio, remove noise, or amplify resonance, emphasize harmonic rich melodies, etc.
• Multiband Phase automatically identifies and fixes phase-related issues
• XSpecMatch is a real-time audio matching equalizer

👉 Project address:

https://cslmusicteam.sony.fr/prototypes/multiband-phase/

0 3

limited time, you draw and I guess the game

🌟 How to play:

Can neural networks learn to recognize graffiti? In this little game, the user draws from a proposition within 20 seconds, and the neural network tries to guess what you are drawing. The more users play, the more it learns, an interesting use case for machine learning.

👉 Play address:

https://quickdraw.withgoogle.com/

🖌️ Graffiti dataset:

https://quickdraw.withgoogle.com/data

Reference:

[1] Waifu Labs：https://waifulabs.com/blog/ai-creativity

[2] Sony's AI exploration: https://ai.sony/projects/gastronomy/

[3.1] Meta AI brings children's drawings to life: https://ai.facebook.com/blog/using-ai-to-bring-childrens-drawings-to-life/

[3.2] Demo address: https://sketch.metademolab.com/

[4] AI colonoscopy-assisted examination: http://www.chinanews.com.cn/cj/2022/01-07/9646539.shtml

[5] Jeff Dean's annual summary: https://ai.googleblog.com/2022/01/google-research-themes-from-2021-and.html

---

With a vision to redefine data science, Zilliz is committed to building a global leader in open source technology innovation and unlocking the hidden value of unstructured data for enterprises through open source and cloud-native solutions.

Zilliz built the Milvus vector database to accelerate the development of a next-generation data platform. The Milvus database is a graduate project of the LF AI & Data Foundation. It can manage a large number of unstructured data sets and has a wide range of applications in new drug discovery, recommendation systems, chatbots, etc.