Research towards AI models that can generalize, scale and accelerate science
Next week marks the beginning of the 11th International Conference on Learning Expressions (ICLR), May 1-5 in Kigali, Rwanda. It will be the first major artificial intelligence (AI) conference to be held in Africa and the first in-person event since the start of the pandemic.
Researchers from around the world gather to share cutting-edge research in deep learning across areas of AI, statistics, data science, and applications such as machine vision, gaming, and robotics. We are proud to support the conference as a Diamond sponsor and his DEI Champion.
The entire DeepMind team has published 23 papers this year. Here are some highlights:
Open questions on the road to AGI
Recent advances have shown AI’s incredible performance on text and images, but more research is needed before the system can be generalized across multiple domains and scales. This is an important step in developing artificial general intelligence (AGI) as a tool to transform everyday life.
We introduce a new approach in which models learn by solving two problems in one. At the same time, he trains the model to look at the problem from two perspectives, so that the model learns how to reason about tasks that need to solve similar problems. This helps generalize. We also explored the ability of neural networks to generalize by comparing them to the Chomsky hierarchy of languages. Rigorous testing of 2200 models on 16 different tasks revealed that certain models struggled to generalize, suggesting that enriching models with external memory could improve performance. I found it important.
Another challenge we are grappling with is how to proceed with long-term tasks at the expert level with little reward. We developed a new approach and an open-source training data set to help our models learn to explore in a human-like way over time.
innovative approach
As we develop more advanced AI capabilities, we need to ensure that our current methods work as intended and efficiently in the real world. For example, language models can come up with impressive answers, but often fail to explain their responses. We show how language models can be used to solve multi-stage inference problems by leveraging their underlying logical structure to provide explanations that humans can understand and confirm. Adversarial attacks, on the other hand, are a way to explore the limits of an AI model by pushing it to create wrong or harmful output. Training on adversarial examples makes the model more robust against attacks, but may sacrifice performance on “regular” inputs. We show that by adding an adapter, we can create a model that allows us to control this trade-off on the fly.
Reinforcement learning (RL) has proven successful for a variety of real-world challenges, but RL algorithms typically perform one task well and generalize to new tasks poorly. Designed to work hard. I propose algorithmic distillation. This is how we train a transformer to mimic the learning history of RL algorithms across a variety of tasks, allowing a single model to generalize efficiently to new tasks. Reinforcement learning models are also trained by trial and error, which can be very data intensive and time consuming. Nearly 80 billion frames of data were required for model agent 57 to reach human-level performance on 57 Atari games. We share a new way to train to this level with 1/200th the experience and drastically reduce your compute and energy costs.
AI for science
AI is a powerful tool for researchers to analyze vast amounts of complex data and make sense of the world around us. Several papers show how AI is accelerating the progress of science and how science is advancing AI.
Predicting molecular properties from 3D structures is important for drug discovery. We present a denoising method that achieves a new state of the art in molecular property prediction, enables large-scale pre-training, and generalizes across diverse biological datasets. We also introduce a new converter that allows more accurate quantum chemical calculations using only atomic position data.
Finally, using FIGnet, we take inspiration from physics to model collisions between complex shapes such as teapots and donuts. This simulator is applicable across robotics, graphics, and mechanical design.
See the full list of DeepMind papers and the schedule of events at ICLR 2023.
