Tests
Creating next-generation AI agents, discovering modern methods and pioneering learning from the ground up
Next week, AI researchers from around the world will gather on January 12 International conference on learning representations (ICLR), which will take place from May 7 to 11 in Vienna, Austria.
Raia Hadsell, vice president of research at Google DeepMind, will deliver a keynote speech summarizing the last 20 years of work in the field and highlighting how the lessons learned are shaping the future of artificial intelligence for the benefit of humanity.
We will also offer live demonstrations showing how we bring our fundamental research to reality, from developing robotic transformers to creating open source toolkits and models such as Donut.
Teams from across Google DeepMind will present over 70 papers this year. Some research highlights:
Problem-solving agents and human-inspired approaches
Immense language models (LLM) are already revolutionizing advanced AI tools, but their full potential remains untapped. For example, LLM-based AI agents capable of taking effective actions could transform digital assistants into more helpful and intuitive AI tools.
AI assistants that follow natural language instructions and perform online tasks on behalf of humans would significantly save time. In the oral presentation we present Network agentan LLM-based agent who learns from first-hand experience how to navigate and manage elaborate tasks on real-world websites.
To further enhance the overall usefulness of the LLM, we have focused on developing their problem-solving skills. We show how we achieved this by equipping an LLM-based system with a traditionally human approach: making and using “tools”. Separately, we present a training technique that ensures more consistent performance of language models socially acceptable outcomes. Our approach uses a sandbox rehearsal space that represents social values.
Pushing the boundaries in vision and coding
Our Animated Scene Transformer (DyST) model uses real-world video footage from a single camera to extract 3D representations of objects in the scene and their movements.
Until recently, gigantic AI models focused primarily on text and images, laying the foundation for large-scale pattern recognition and data interpretation. The field is now moving beyond these stationary realms to encompass the dynamics of real-world visual environments. As IT advances across the board, it becomes increasingly crucial to generate and optimize source code with maximum efficiency.
When you watch video on a flat screen, you intuitively understand the three-dimensional nature of the scene. Machines, however, have difficulty imitating this ability without explicit supervision. We present ours Dynamic scene transformer (DyST), which uses real video footage from a single camera to extract three-dimensional representations of objects in a scene and their movements. Moreover, DyST also allows you to generate novel versions of the same video, with user control over camera angles and content.
Emulating human cognitive strategies also allows for better AI code generators. When programmers write elaborate code, they typically “break down” the task into simpler subtasks. WITH ExeDecwe introduce a novel code generation approach that uses a decomposition approach to boost the efficiency of programming and generalization of AI systems.
In parallel reflector paper We explore the novel employ of machine learning not only for code generation but also for code optimization, introducing: a dataset enabling robust benchmarking of code performance. Code optimization is challenging and requires elaborate reasoning, and our dataset allows exploration of a range of machine learning techniques. We show that the resulting learning strategies outperform human code optimizations.
ExeDec introduces a novel code generation approach that uses a decomposition approach to improve the efficiency of programming and generalization of AI systems
Progress in primary education
Our research teams address the most crucial issues in AI – from exploring the nature of machine cognition to understanding how advanced AI models generalize – while working to overcome key theoretical challenges.
For both humans and machines, causal inference and the ability to predict events are closely related concepts. In the spotlight presentation, we look at how to do this Reinforcement learning is influenced by prediction-based training goalsand draw parallels to changes in brain activity also related to prediction.
When AI agents can generalize well to modern scenarios, is it because, like humans, they have learned a basic causal model of their world? This is a key question in advanced artificial intelligence. In the oral presentation we reveal that such models we have indeed learned an approximate causal model the processes that led to the training data and discuss their profound implications.
Another key issue in AI is trust, which depends in part on how accurately models can estimate the uncertainty of results – a key factor in reliable decision-making. We did significant advances in uncertainty estimation within deep Bayesian learningusing a plain and essentially cost-free method.
Finally, we study game theory’s Nash equilibrium (NE) – a state in which no player benefits from changing his strategy if others maintain theirs. Outside of plain two-player games, even approximating a Nash equilibrium is computationally complex, but in oral presentation reveal new, cutting-edge approaches in negotiating offers from poker to auctions.
Connecting the AI community
We are delighted to sponsor ICLR and support initiatives including Queer in AI AND Women in machine learning. Such partnerships not only strengthen research collaborations, but also support a dynamic, diverse AI and machine learning community.
If you are at ICLR, be sure to visit our booth and ours Google Research friends from the neighborhood. Discover our pioneering research, meet our workshop teams and connect with our experts at the conference. We look forward to hearing from you!