Radical Reads

The International Conference on Machine Learning (ICML) is one of the premier events in the field of machine learning. This year, the conference received over 9,400 paper submissions, with 2,610 selected for presentation. Among these, only ten were honoured with a best paper award. Notably, many of this year’s awarded papers emerged from collaborations between industry and academia, highlighting the increasing prevalence of such partnerships. In 2023, 21 significant machine learning models resulted from these types of collaborations, a sharp rise from just one major contribution in 2022. This week we highlight three of the awarded papers which reflect emerging research trends.

Debating with More Persuasive LLMs Leads to More Truthful Answers

A study by cross-industry and academic researchers including University College London, Anthropic, FAR AI, and Speechmatics shows that AI models debating each other can lead to more accurate answers. By having AI argue for different answers, both non-expert models and humans are better at identifying the correct one.

Key Contributions:

• Enhanced Accuracy: AI debates significantly improve answer accuracy compared to non-adversarial methods.
• Persuasiveness Matters: Optimizing AI for persuasiveness increases argument accuracy, using a metric that does not rely on predefined answers.
• Human Oversight: Human judges outperform AI in assessing debates, showing the value of human involvement.
• Scalable Method: Debates could be a scalable way to oversee increasingly advanced AI models.

This research highlights that as AI evolves, optimizing for persuasiveness in debates could result in more reliable information.

Considerations for Differentially Private Learning with Large-Scale Public Pretraining

Researchers from ETH Zürich, the University of Waterloo, and Google DeepMind highlight the risks to confidentiality in current training paradigms for large language models (LLMs). The authors argue that there is a need to address privacy challenges in machine learning, particularly when models are trained on large public datasets. 

Key Contributions:

• Critique of Training Paradigms: The authors argue that current methods compromise confidentiality for two main reasons:
  • Overestimated Value of Public Pre-Training: Public and private data distributions often overlap, leading to a false sense of security.
  • Outsourcing of Private Data: The significant computational power required for training large models necessitates outsourcing, risking data privacy.

The paper calls on the research community to develop solutions to these privacy concerns, especially as smaller models gain popularity.

Genie: Generative Interactive Environments
Researchers from Google DeepMind and the University of British Columbia have introduced Genie, a generative interactive environment trained unsupervised using unlabelled Internet videos. Genie can generate various action-controllable virtual worlds based on text descriptions, synthetic images, photos, or sketches.

Key Contributions:

• Scalable Training: By training Genie to predict actions between video game frames without supervision, it can efficiently scale data during training.
• Next State Prediction: The model can also predict the next state by analyzing past states and user input actions.
• Behaviour Imitation from Unseen Videos: Genie uses spatiotemporal transformers, a novel video tokenizer, and a causal action model to extract latent actions and create dynamic content. These latent actions, combined with video tokens, allow the dynamics model to autonomously predict future frames.

The paper reflects a growing interest in advancing beyond text-centric LLMs toward multimodal models that integrate physical and digital experiences. Recently, Stanford Professor and Radical Ventures Scientific Partner Fei-Fei Li, shared her vision of building spatial intelligence with AI – enabling systems to process visual data, make predictions, and interact with humans in the real world.