The latest AI news from last week
From Transformers in Robotics (Google AI) to multimodal SSL (Meta AI) to requirements engineering in ML (CMU) to free courses. This week is full of fascinating announcements: research papers, technologies behind popular services, and open-source announcements!
Here is an outline of the news. We cover main points of each of them in the article below.
ML Research
Google AI presented RT-1: Robotics Transformer model 😮
Meta AI published a paper detailing the data2vec 2.0 algorithm with an increased training efficiency (a new version of data2vec)
OpenAI presents an improved embedding model priced 99.8% lower than the Davinci model
Researchers from Carnegie Mellon University presented a framework for requirements engineering in machine learning
ML Applications
Amazon Science detailed the methods behind continual learning in Alexa
Google AI thought Recorder to detect different speakers in real time
Meta AI open-sourced its Anonymous Credential Service (ACS) for user authentification in a de-identified manner
Announcements
Hugging Face created the “Ethics and Society Newsletter” and already published its #2 edition
Machine Learning University curated by Amazon Science presented their free, online Responsible AI course
IBM released the first episode of its free series explaining how quantum computing works
Let’s dive into the news! Make sure you read the article to find all links.
ML Research
Google AI presented RT-1: Robotics Transformer model 😮
New research from Google AI addresses the lack of efficient, scalable, and fast-enough-for-real-time-inference models in robotics.
Robotics Transformer 1 (RT-1) is a multi-task model that tokenizes robot inputs and outputs actions (e.g., camera images, task instructions, and motor commands) to enable efficient inference at runtime, which makes real-time control feasible.
Links: original blog post, original paper, the project’s website
Meta AI published a paper detailing the data2vec 2.0 algorithm with an increased training efficiency (a new version of data2vec)
data2vec 2.0 performs as well as these models with significant improvements in pre-training time:
Masked Autoencoders in 16.4x less time on computer vision
wav2vec 2.0 in 10.6x less time on Librispeech speech recognition
RoBERTa model in 2x less time on GLUE natural language understanding
Links: original blog post, original paper, models and code here
OpenAI presents an improved embedding model priced 99.8% lower than the Davinci model
The embedding model converts a text concept into a numeric vector which then can be used for various ML tasks. This is a numeric representation of the text that
allows words with similar meanings to have a similar representation. From Machine Learning Mastery post.
The new model OpenAI model is called text-embedding-ada-002. It outperforms all the old embedding models on these kinds of tasks:
text search
code search
sentence similarity
It gets comparable performance on text classification.
At the same time, it costs 99.8% lower than the Davinci model, the previous most capable OpenAI model.
Links: original blog post, documentation, pricing
Researchers from Carnegie Mellon University presented a framework for requirements engineering in machine learning
Requirements engineering (RE) is a crucial component of software engineering. When it comes to ML, this practice is not standardized and properly defined and considered one of the hardest tasks in ML development.
To bridge this gap, researchers created a simple evaluation framework.
Link: original blog post
ML Applications
Amazon Science detailed the methods behind continual learning in Alexa
The company created a blog post describing the approach allowing Alexa to detect misheard or misspelled queries and self-correct them.
The concept is called Query Rewriting (QR) when the conversational AI agent self-detects and self-corrects incorrect queries. Amazon Science presented two papers:
The first paper addresses a query rewriting problem by a novel Constrained Generation Framework (CGF)
The second paper addresses a limitation of the rewrite approach
Google AI thought Recorder to detect different speakers in real time
Recorder is not just an audio-recording app. It leverages machine learning to provide more capabilities for the user. For example, it can transcribe speech or suggest tags for titles, etc.
Last week, Google AI published an article about the research behind Recorder’s new capability — the “speaker labels” feature.
The system mainly consists of three components: a speaker turn detection model that detects a change of speaker in the input speech, a speaker encoder model that extracts voice characteristics from each speaker turn, and a multi-stage clustering algorithm that annotates speaker labels to each speaker turn in a highly efficient way. From the original blog post
Links: original blog post, paper, open-sourced spectral clustering algorithm here
Meta AI open-sourced its Anonymous Credential Service (ACS) for user authentification in a de-identified manner
Autentification in a de-identified manner is part of a Meta strategy towards data minimization. The article details the approach behind Anonymous Credential Service (ACS).
At a high level, ACS supports de-identified authentication by splitting authentication into two phases, token issuance and token redemption. From the original blog post
Announcements
Hugging Face created the “Ethics and Society Newsletter” and already published its #2 edition
Links: Edition #1, edition #2
Machine Learning University curated by Amazon Science presented their free, online Responsible AI course
Links: original blog post, YouTube playlist of the course
IBM released the first episode of its free series explaining how quantum computing works
Links: original blog post, YouTube playlist of the course, textbook
If you liked this edition, make sure to 📨 share the article with your friends or 💭 write a post in social media. I plan to do these weekly digests regularly and your reactions are important! Each of your actions shows me your support! 🙌🏼