Focus AI #110

#Google #RT1 #Baidu # Stanford #EarSPy

Focus AI #110, the latest AI news in 3 minutes

Happy new year! Let’s get started!!! 🔥

ps: if you like this content, feel free to share it or to contribute with a small (or large) gift subscription)

This week:

• 🙉 EarSpy can eavesdrop on your phone conversations using motion sensors

• 🤖 Google trains a big model to create smart robots

• 👩‍💻 Baidu makes a multilingual coding assistant

• 🎼 Generate music from spectrograms via StableDiffusion

• 💉 PubMed GPT, an AI model trained to interpret biomedical language

And about code:

• ⚒️ Top Python libraries of 2022

• 📊 NumPy 1.24 is out

• 🐍 Classifying Python virtual environment workflows

== News ==

🙉 EarSpy can eavesdrop on your phone conversations using motion sensors

Researchers have developed an attack that captures what people say through their phone by gathering vibrations from the phone's loudspeaker.

The EarSpy attack was tested by playing voice samples and analyzing the accelerometer data with neural network tools. The EarSpy attack was able to identify the gender of the speaker and the words spoken with high accuracy, and could detect the person's identity with an accuracy rate of 91.24%.

The researchers recommend that smartphone makers position motion sensors away from sources of vibrations and reduce sound pressure during phone calls to remedy this vulnerability.

🤖 Google trains a big model to create smart robots

The Robotics Transformer model, RT-1, is a machine learning model for robots that absorbs large amounts of diverse data to enable it to generalize to new tasks, environments, and objects.

RT-1 was trained on a large dataset of multi-task demonstrations and is able to absorb data from different domains such as simulation or different robots.

RT-1 was tested on real robots and showed promising results in terms of its ability to generalize and adapt to new tasks, environments, and objects.

👩‍💻 Baidu makes a multilingual coding assistant

Baidu has created ERNIE-Code, a 560 million parameter coding model that is pre-trained on six programming languages and over 100 languages via the CommonCrawl-100 corpus.

ERNIE-Code is able to perform competitively on tasks such as code summarization, code generation, document translation, and program repair, and performs significantly better on code summarization compared to other models.

The creation of ERNIE-Code is intended to mitigate the English-centric bias in program pre-training and increase the representation of other languages in AI.

🎼 Generate music from spectrograms via StableDiffusion

You've heard of Stable Diffusion, the open-source AI model that generates images from text? Well, they fine-tuned the model to generate images of spectrograms. The magic is that spectrograms can then be converted to an audio clip.

This is the v1.5 stable diffusion model with no modifications, just fine-tuned on images of spectrograms paired with text. Audio processing happens downstream of the model.

It can generate infinite variations of a prompt by varying the seed. All the same web UIs and techniques like img2img, inpainting, negative prompts, and interpolation work out of the box.

💉 PubMed GPT, an AI model trained to interpret biomedical language

A partnership between MosaicML and the Stanford Center for Research on Foundation Models (CRFM) has released PubMed GPT, an artificial intelligence model that interprets biomedical language.

Trained on the MosaicML Cloud platform, the 2.7 billion parameter GPT was built using the Pile dataset of 16 million abstracts and five million full-text articles from the biomedical literature, and it has achieved state-of-the-art results on medical question and answer text from the US Medical Licensing Exam (USMLE).

The model was designed to demonstrate the capabilities of domain-specific large language models (LLMs) and the power of off-the-shelf LLM training recipes.

== Code & Tools ==

⚒️ Top Python libraries of 2022

Long story, short: Ruff (linter), python-benedict, Memray, Codon, LangChain, fugue, Diffusers, LineaPy, whylogs, Mito

📊 NumPy 1.24 is out

The NumPy 1.24.0 release continues the ongoing work to improve the handling and promotion of dtypes, increase the execution speed, and clarify the documentation. There are also a large number of new and expired deprecations due to changes in promotion and cleanups. This might be called a deprecation release.

🐍 Classifying Python virtual environment workflows

The way people manage their virtual environments in Python can be classified into three categories: tools that fully manage the virtual environment life cycle, virtual environment helpers, and manual virtual environment management. Virtual environments can also be stored locally in the workspace directory or in a central directory for the user. The method for managing and storing virtual environments can depend on the project and personal preferences of the user.

#Google #RT1 #Baidu #Stanford #EarSPy

A big thanks to our sources: https://jack-clark.net/, https://www.actuia.com/, https://thevariable.com/news/, https://techcrunch.com/, https://read.deeplearning.ai/the-batch/

What about you? Have you noticed something else?

Don’t miss the next news!

Have a good week-end,

Maxime 🙃 from Toulouse, France with 🌺