DeepMind’s Path to Artificial General Intelligence
Introduction
DeepMind Alberta recently released a new paper titled “The Alberta Plan” which discusses their objectives over the next decade. The paper mainly looks at Reinforcement Learning (RL) and briefly touches upon other topics. What’s interesting is its contrarian views as it approaches different problems.
Their main goal is to achieve better continual learning in the world.
Agents will not only learn but learn to learn (meta-learning) to achieve their goals.
Analysis
Machine learning (ML) engineers currently fine-tune models to perform efficiently. This is labour-intensive as it requires hand-selecting datasets, utilizing subject matter experts, and integrating human knowledge. As datasets scale these fine tuned models will need more human intervention and can’t adapt on their own. DeepMind wants to create algorithms which adapt as the dataset evolves without more human involvement. This approach will be challenging as it goes against the current norm. However, it should result in algorithms that can have multiple use cases.
Algorithms often change learning rates to learn more efficiently. They do this by placing weights on different rewards or slowing things down once functions converge. The Alberta Plan believes in treating every point in time the same. When algorithms begin to fail, they should be able to adapt to different environments. Environments change over time; if temporal uniformity can be assumed, then algorithms should be able to adapt better.
Compute power is increasing yearly and believers in the scaling hypothesis think that our current models are fine and we just need compute power to catch up and make models larger. While you may doubt this, it is undeniable that compute power is a large determinant of the performance of algorithms. As such, the paper suggests focusing on methods that scale well with compute power and spend less attention on things that are harder to scale, such as the human labelling of data.
Multi-agent environments are significantly more effective at learning than single agents. Agents respond to decisions made by each other resulting in cooperative learning. The paper discusses something known as intelligence amplification, where human performance is enhanced by software and onwards. The goal is to make this effective in software where humans are unnecessary. Agents should simply be amplified by interacting with each other. The team at DeepMind believes this is pivotal in unlocking the power of AI.
Conclusion
Current models are tailor-made to solve specific problems to maximize performance and do well on some benchmarks. We are working with things that, unfortunately, don’t scale well, and this different perspective could be what we need to create algorithms that can solve various problems. Moving in this direction could be the best way to reach some form of Artificial General Intelligence (AGI).
Reference: https://arxiv.org/pdf/2208.11173v2.pdf
