Decoding speech from brain activity ...monthly round-up of #AI news

Happy October. Last month in our inaugural newsletter we discovered how #artificialintelligence?helped early detection of dreaded diseases. Let's discover what's in the AI platter this time.

In this edition we cover:

• How AI is used to decode speech from brain activity.
• DeepMind AI invents?faster computation algorithm

How AI is used to decode speech from brain activity

Decoding speech from brain activity is yet another milestone that has found its way through AI. In the past, the progress was largely through invasive brain recording techniques like stereotactic electroencephalography?and?electrocardiography. Studies have revealed that non-invasive methods provide a safe and scalable solution. This however is quite challenging as the non-invasive methods are noisy and complex.

To address this, an AI solution was developed by a team of researchers at Meta which aligns brain signals to speech segments.

Research Credits: Alexandre Défossez,?Charlotte Caucheteux,?Jérémy Rapin,?Ori Kabeli, and Jean-Rémi King

The team addressed the challenges by building a deep learning model trained with contrastive learning and an open-source, self-supervised learning model - wave2vec 2.0 pre-trained on 56k hours of speech to align speech sounds and brain recordings.

The Model was evaluated on four public datasets containing the brain activity of 169 participants while listening to natural speech like audiobooks etc and the accurate response was in the AI’s top ten guesses up to seventy-three percent of the time.

Research Paper

Meta AI Blog

DeepMind AI invents?faster computation algorithms for Math Puzzles

Researchers at Deep mind London have developed a Deep Reinforcement learning based on AlphaZero as well as a game-playing method called Tree search to solve tough mathematical calculations and Matrix Multiplication with improved computing efficiency. The agent AlphaTensor has been trained to play a single-player game to efficiently predict tensor decompositions. DeepMind’s approach was published in The Nature.

The researchers performed testing on matrices up to 5 × 5. Alpha tensor performed exceptionally well in some cases like when multiplying a 4 × 5 matrix by a 5 × 5 matrix the algorithm needed only 76 individual multiplications whereas the previous best was 80 individual multiplications.

To handle larger matrix multiplications the researchers created a meta-algorithm that divided the problem into smaller ones. When crossing an 11 × 12 and a 12 × 12 matrix, their method reduced the number of required multiplications from 1,022 to 990.

News making waves #artificialintelligence

Meta’s Make-A-Video: a system built by Uriel Singer and colleagues at Meta, turns text prompts into high-resolution video clips. Earlier this year Meta announced Make-A-Scene, a multimodal generative AI method,?where people can create photorealistic illustrations using texts.

DALLE-2 waitlist removed: Open AI's DALL·E beta has removed the waitlist for all. New users can start creating straight away. They are also testing DALL·E API with several customers?and can soon offer it to developers and businesses so they can build apps on this powerful?system

A DALL.E generated Image from the text "An oil painting of?sunrise in Van Gogh style"

Add to your AI Playlist: Andrew NG's Latest Ted Talk: "How AI could empower any business" and OpenAI CEO Sam Altman's Interview with Reid Hoffman: AI for the next Era.

That's all for now, stay connected via?Youtube?and?Twitter?to get updates from Data Science and the?#machinelearning?world.

Happy Learning!