Back
2025INF020.pdf
Doctoral thesis

Automating AI research

  • 2025

PhD: Università della Svizzera italiana

General-purpose meta learning
Recursive self-improvement
AI scientist
Self-referential learning
In-context learning
Large language models
English A core element of Artificial General Intelligence (AGI) and ultimately Artificial Superintelligence (ASI) is the ability of AI to improve itself, including its own learning algorithm. This single capability, once developed by human researchers, could initiate an autonomously self-improving system capable of independently conducting all subsequent research. Toward this vision, this thesis advances the automation of AI research by developing methods for AI systems to discover general-purpose learning algorithms. Today, machine learning (ML) research still relies heavily on human-designed learning algorithms, architectures, losses, optimizers, data, and other components. This process requires significant manual effort and the selection of appropriate inductive biases, limited by human creativity and knowledge. Meta-learning, or learning-to-learn, instead aims to automate the process of artificial intelligence (AI) research and promises to unlock greater capabilities with less manual effort. In recent years, meta-learning has made significant progress in producing models that can learn very quickly from a few examples using in-context learning, fast-weights, and optimization-based methods. This tremendously helps in adapting to new, similar problems, but it does not automate AI research itself. To address this limitation, we introduce meta-learners that discover general-purpose learning algorithms. The goal is the discovery of novel learning algorithms that can be reused across a wide range of tasks, similar to human-engineered learning algorithms. We present several novel methods addressing meta-generalization, including learned loss functions, weight-shared LSTMs that implement gradient descent in their recurrent dynamics, and black-box Transformers that learn how to in-context learn generally. The main contributions include MetaGenRL, a novel off-policy gradient-based meta-RL algorithm that, for the first time, discovers general-purpose learning algorithms with high performance on diverse robotic control tasks. Our Variable Shared Meta-Learners (VSML) make use of parameter sharing and sparsity in meta-learning to automatically discover powerful in-context learning algorithms that neither require explicit gradients at meta-test time nor weight updates. Next, we introduce SymLA, a method that builds on the neural network symmetries of VSML to improve the generalization of learning algorithms in meta-RL. In our general-purpose in-context learners (GPICL), we discover phase transitions where models transition from memorization to task identification, to general learning-to-learn. We identify crucial ingredients such as memory capacity and practical interventions in meta-training. We explore how these insights can be applied to meta-RL in generally learning agents (GLAs). Our results show that meta-learning is a powerful approach for generating general-purpose learning algorithms that can be effectively transferred to new and unseen environments. Finally, we seek AI to self-improve while minimizing its dependence on human engineering. To this end, we explore self-referential systems that can recursively improve themselves without hard-wired meta-optimization. We extend this idea to AI scientists, large language models that can automate AI research by generating hypotheses, conducting experiments, and interpreting results.
Collections
Language
  • English
Classification
Computer science and technology
License
License undefined
Open access status
green
Identifiers
Persistent URL
https://n2t.net/ark:/12658/srd1334926
Statistics
Document views: 0 File downloads:
  • 2025INF020.pdf: 0