We’re back with more suggestions from our researchers for ways to expand your knowledge of AI.
Today’s #AtHomeWithAI recommendations are from research scientist Kimberly Stachenfeld (@neuro_kim) (1/7)
She recommends “The Scientist in the Crib” [longer listen] by @AlisonGopnik, Andrew Meltzoff, & Patricia K. Kuhl for those who are interested in what early learning tells us about the mind.
Interested in computational systems neuroscience? @neuro_kim recommends the lecture series from @MBLScience to learn more about circuits and system properties of the brain.
@neuro_kim says Theoretical Neuroscience: Computational and Mathematical Modeling of Neural Systems [longer read] by Peter Dayan & L.F. Abbott is a must read for anyone looking for an introduction to the topic.
Described as “a classic for anyone who wants to understand the roots of DL”, Kimberly recommends “The Appeal of Parallel Distributed Processing” [longer read] by James McClelland, the late David Rumelhart, & Geoffrey Hinton.
Our latest generative technology is now powering MusicFX DJ in @LabsDotGoogle - and we’ve also updated Music AI Sandbox, a suite of experimental music tools which can streamline creation. 🎵
This will make it easier than ever to make music in real-time with AI. ✨goo.gle/4eTg28Z
MusicFX DJ lets you input multiple prompts and include details on instruments, genres and vibes to create music. 🎛️
We’ve updated and improved the interface using feedback from @YouTube’s Music AI Incubator.
Two key innovations lie at the core of MusicFX DJ.
🔘 We adapted our models to perform real-time streaming by training them to generate the next 2 seconds of music, based on the previous 10 seconds.
🔘 “Style embedding” is steered by the player, which is a mix of text prompt embeddings set by the slider values
Meet our AI-powered robot that’s ready to play table tennis. 🤖🏓
It’s the first agent to achieve amateur human level performance in this sport. Here’s how it works. 🧵
Robotic table tennis has served as a benchmark for this type of research since the 1980s.
The robot has to be good at low level skills, such as returning the ball, as well as high level skills, like strategizing and long-term planning to achieve a goal.
To train the robot, we gathered a dataset of initial table tennis ball states - which included information about position, speed, and spin.
The system practiced using this library and learned different skills, like forehand topspin, backhand targeting, and returning serves.
AI systems can be powerful but opaque "black boxes" - even to researchers who train them. ⬛
Enter Gemma Scope: a set of open tools made up of sparse autoencoders to help decode the inner workings of Gemma 2 models, and better address safety issues. → dpmd.ai/gemma-scope
Language models turn your text input into a series of ‘activations’ - which map the relationships between the words you’ve entered to help it write its answer. 💬
Activations at different layers in its neural network represent increasingly advanced concepts, known as ‘features’.
Activations are made up of neurons, which “fire” for many unrelated features - making them hard to decipher.
Each feature seems to be a specific combination of neurons - but how can we find the meaningful combinations of neurons?
We’re also introducing ShieldGemma: a series of state-of-the-art safety classifiers designed to filter harmful content. 🛡️
These target hate speech, harassment, sexually explicit material and more, both in the input and output stages.
Finally, we’re announcing Gemma Scope, a set of tools to help researchers examine how Gemma 2 makes decisions. 🔍
It's a comprehensive, open suite of sparse autoencoders - specialized neural networks that zoom into the model’s inner workings and make them more interpretable.
We’re presenting the first AI to solve International Mathematical Olympiad problems at a silver medalist level.🥈
It combines AlphaProof, a new breakthrough model for formal reasoning, and AlphaGeometry 2, an improved version of our previous system. 🧵 dpmd.ai/imo-silver
Our system had to solve this year's six IMO problems, involving algebra, combinatorics, geometry & number theory. We then invited mathematicians @wtgowers and Dr Joseph K Myers to oversee scoring.
It solved 4️⃣ problems to gain 28 points - equivalent to earning a silver medal. ↓
For non-geometry, it uses AlphaProof, which can create proofs in Lean. 🧮
It couples a pre-trained language model with the AlphaZero reinforcement learning algorithm, which previously taught itself to master games like chess, shogi and Go. dpmd.ai/imo-silver