**Paving the way for advanced AI systems at ICML 2022**
The thirty-ninth International Conference on Machine Learning (ICML 2022) is taking place from 17-23 July, 2022 at the Baltimore Convention Center in Maryland, USA. This conference brings together researchers from various fields to share their groundbreaking work in machine learning.
At ICML 2022, our research teams are presenting 30 papers, including 17 collaborations with external partners. We are also proud to sponsor the conference and support the workshops and social events organized by LatinX, Black in AI, Queer in AI, and Women in Machine Learning. Here’s a glimpse of our upcoming oral and spotlight presentations:
**Improving reinforcement learning**
To develop more advanced AI systems, it is essential to enhance reinforcement learning (RL) algorithms. We are presenting a new approach called generalised policy improvement (GPI), which boosts an agent’s performance by applying it to compositions of policies. Another interesting presentation explores an efficient way of exploration without the need for bonuses. Additionally, we propose a method to augment RL agents with a memory-based retrieval process, enabling them to utilize past experiences effectively.
**Advancements in language models**
Language plays a crucial role in human communication and understanding. To improve language models, we are exploring unified scaling laws and retrieval techniques. Additionally, we have developed a new dataset called StreamingQA, which evaluates how models adapt to and forget new knowledge over time. Our research also focuses on narrative generation, where we address the challenge of maintaining entity coherence and consistency in current pretrained language models, which struggle with longer texts due to short-term memory limitations.
**Algorithmic reasoning**
The field of neural algorithmic reasoning shows promise in adapting known algorithms to real-world problems. We are introducing the CLRS benchmark, which evaluates neural networks’ performance on thirty classical algorithms. Additionally, we propose a general incremental learning algorithm that applies hindsight experience replay to automated theorem proving. We also present a framework for constraint-based learned simulation, which combines traditional simulation and numerical methods in machine learning simulators for complex simulation problems in science and engineering.
To learn more about our work at ICML 2022, visit our event page [here](https://deepmind.events/events/icml2022).
By attending and presenting at ICML 2022, we aim to advance the field of AI and pave the way for more effective and efficient AI systems in the future.