Google DeepMind at ICML 2024

Exploring AGI, the challenges of scaling and the way forward for multimodal generative AI

Subsequent week the substitute intelligence (AI) neighborhood will come collectively for the 2024 Worldwide Convention on Machine Studying (ICML). Operating from July 21-27 in Vienna, Austria, the convention is a world platform for showcasing the newest advances, exchanging concepts and shaping the way forward for AI analysis.

This yr, groups from throughout Google DeepMind will current greater than 80 analysis papers. At our sales space, we’ll additionally showcase our multimodal on-device mannequin, Gemini Nano, our new household of AI fashions for schooling referred to as LearnLM and we’ll demo TacticAI, an AI assistant that may assist with soccer ways.

Right here we introduce a few of our oral, highlight and poster shows:

Defining the trail to AGI

What’s synthetic common intelligence (AGI)? The phrase describes an AI system that’s a minimum of as succesful as a human at most duties. As AI fashions proceed to advance, defining what AGI may appear to be in follow will change into more and more vital.

We’ll current a framework for classifying the capabilities and behaviors of AGI fashions. Relying on their efficiency, generality and autonomy, our paper categorizes techniques starting from non-AI calculators to rising AI fashions and different novel applied sciences.

We’ll additionally present that open-endedness is crucial to constructing generalized AI that goes past human capabilities. Whereas many current AI advances have been pushed by present Web-scale information, open-ended techniques can generate new discoveries that stretch human data.

Scaling AI techniques effectively and responsibly

Creating bigger, extra succesful AI fashions requires extra environment friendly coaching strategies, nearer alignment with human preferences and higher privateness safeguards.

We’ll present how utilizing classification as an alternative of regression strategies makes it simpler to scale deep reinforcement studying techniques and obtain state-of-the-art efficiency throughout totally different domains. Moreover, we suggest a novel strategy that predicts the distribution of penalties of a reinforcement studying agent’s actions, serving to quickly consider new eventualities.

Our researchers current an alignment-maintaining strategy that reduces the necessity for human oversight, and a new strategy to fine-tuning massive language fashions (LLMs), primarily based on recreation concept, higher aligns a LLM’s output with human preferences.

We critique the strategy of coaching fashions on public information and solely fine-tuning with “differentially personal” coaching, and argue this strategy could not supply the privateness or utility that’s typically claimed it does.

New approaches in generative AI and multimodality

Generative AI applied sciences and multimodal capabilities are increasing the inventive potentialities of digital media.

We’ll current VideoPoet, which makes use of an LLM to generate state-of-the-art video and audio from multimodal inputs together with pictures, textual content, audio and different video.

And share Genie (generative interactive environments), which might generate a spread of playable environments for coaching AI brokers, primarily based on textual content prompts, pictures, pictures, or sketches.

Lastly, we introduce MagicLens, a novel picture retrieval system that makes use of textual content directions to retrieve pictures with richer relations past visible similarity.

Supporting the AI neighborhood

We’re proud to sponsor ICML and foster a various neighborhood in AI and machine studying by supporting initiatives led by Incapacity in AI,
Queer in AI,
LatinX in AI and
Ladies in Machine Studying.

In case you’re on the convention, go to the Google DeepMind and Google Analysis cubicles to satisfy our groups, see dwell demos and discover out extra about our analysis.