Google DeepMind at ICML 2024

Defining the trail to AGI

What’s Synthetic Normal Intelligence (AGI)? The time period describes an AI system that’s at the very least as succesful as a human in most duties. As AI fashions evolve, it turns into more and more necessary to outline what AGI would possibly appear to be in apply.

We current a framework for classifying the capabilities and behaviors of AGI fashions. Relying on their efficiency, generality and autonomy, our article categorizes programs starting from non-AI computer systems to new AI fashions and different novel applied sciences.

We will even present that openness is crucial to constructing common AI that goes past human capabilities. Whereas many current AI advances have relied on present internet-scale knowledge, open programs can produce new discoveries that advance human data.

Scale AI programs effectively and responsibly

Growing bigger, extra highly effective AI fashions requires extra environment friendly coaching strategies, a more in-depth alignment with human preferences, and higher knowledge safety measures.

We present how utilizing classification methods as a substitute of regression methods makes it simpler to scale deep reinforcement studying programs and obtain state-of-the-art efficiency in varied domains. Moreover, we suggest a novel strategy that predicts the distribution of penalties of a reinforcement studying agent's actions, serving to to shortly consider new eventualities.

Our researchers current an strategy to sustaining alignment that reduces the necessity for human oversight, and a brand new strategy to fine-tuning massive language fashions (LLMs) based mostly on sport idea that higher tailors the output of an LLM to human preferences.

We criticize the strategy of coaching fashions on public knowledge and fine-tuning solely with “differentially personal” coaching, arguing that this strategy could not present the privateness or utility usually claimed.

New approaches in generative AI and multimodality

Generative AI applied sciences and multimodal capabilities broaden the artistic potentialities of digital media.

Introducing VideoPoet, which makes use of an LLM to generate state-of-the-art video and audio knowledge from multimodal inputs corresponding to photos, textual content, audio, and different movies.

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

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

Supporting the AI ​​neighborhood

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

For those who're attending the convention, go to the Google DeepMind and Google Analysis cubicles to fulfill our groups, watch dwell demos, and be taught extra about our analysis.