# Building Machines That Learn and Think Like People

**Authors:** Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, Samuel J. Gershman  
**arXiv:** [1604.00289](https://arxiv.org/abs/1604.00289)  
**Date:** 2016-04-01  
**Summary page:** https://www.summarizepaper.com/en/paper/1604.00289v1/

## Key points

- Recent progress in AI has sparked interest in developing machines that can learn and think like humans
- Deep neural networks have been used to train machines in various tasks, achieving remarkable performance
- AI systems still differ significantly from human intelligence despite their achievements and biological inspiration
- To create truly human-like machines, it is necessary to go beyond current engineering trends
- Machines should build causal models of the world to support explanation and understanding
- Grounding learning in intuitive theories of physics and psychology enhances overall knowledge base
- Harnessing compositionality and learning-to-learn techniques enables rapid acquisition and generalization of knowledge
- Combining recent advances in neural network technology with structured cognitive models can leverage strengths of both approaches.

## Summary

Recent progress in Artificial Intelligence (AI) has sparked a renewed interest in developing machines that can learn and think like humans. This progress has been driven by the use of deep neural networks, which have been trained end-to-end in various tasks such as object recognition, video games, and board games. These systems have achieved remarkable performance, sometimes even surpassing human capabilities. However, despite their impressive achievements and biological inspiration, these AI systems still differ significantly from human intelligence. In order to create truly human-like learning and thinking machines, it is necessary to go beyond current engineering trends in both what they learn and how they learn it. To achieve this goal, the authors propose several key principles that should guide the development of these machines. Firstly, they argue that these machines should be able to build causal models of the world that support explanation and understanding rather than just solving pattern recognition problems. This means going beyond recognizing patterns and instead focusing on understanding the underlying causes and mechanisms behind them. Secondly, the authors suggest grounding learning in intuitive theories of physics and psychology. By incorporating knowledge from these domains into the learning process, machines can acquire a richer understanding of the world and enhance their overall knowledge base. Lastly, the authors emphasize the importance of harnessing compositionality and learning-to-learn techniques to enable rapid acquisition and generalization of knowledge to new tasks and situations. By leveraging these abilities, machines can quickly adapt to new challenges without requiring extensive retraining or explicit programming. The authors conclude by highlighting concrete challenges and promising routes towards achieving these goals. They propose combining recent advances in neural network technology with more structured cognitive models to leverage the strengths of both approaches. In summary, recent advancements in AI have demonstrated impressive performance in tasks such as object recognition and game playing. However, creating truly human-like learning and thinking machines requires going beyond current engineering trends. By building causal models of the world, grounding learning in intuitive theories, harnessing compositionality and learning-to-learn techniques; machines can acquire a deeper understanding of the world and rapidly adapt to new challenges.