Richard Sutton: reinforcement learning and continuous training
This article is a lecture by Richard Sutton on the direction of AI research, emphasizing the importance of understanding how the brain works and building transformation models for progress. He believes that previous research has strayed from the correct solution path, with backpropagation algorithms only solving current problem features while lacking a mechanism to drive learning systems to find generalizing features.
Sutton also discussed reward hypothesis and intelligent prediction, believing the reward hypothesis can be well understood as maximizing a single external scalar signal. He predicts a 25% probability of understanding intelligence by 2030, which would change societal status quo, help us deeper understand thought processes, and trigger technological and economic changes.
Finally, he mentioned brain-computer interface technology, thinking it is an interface but possibly other methods exist for compressing information, finding more efficient communication ways beyond language. He also provided advice to AI researchers: maintain regular writing habits, remain neutral on trends, and choose problems that are both important and likely to yield research results.
This article brought some insights and thoughts from Richard Sutton’s view on AI research direction, hoping to inspire others in exploring artificial intelligence.
Translation
这篇文章是理查德·萨顿对AI研究方向的一次演讲，他强调了理解大脑如何运作和构建转换模型对于进步的重要性。 他认为，之前的研究已经偏离了正确的解决方向，反向传播算法只能解决当前问题的特征，而缺乏驱动学习系统找到泛化特征的机制。
萨顿还谈到了奖励假设和智能预测，他认为奖励假设可以很好地理解为对一个单一外部接收标量信号进行最大化。他预测到2030年，我们有25%的概率能理解智能，而这将会改变社会的现状，帮助我们更深入的认识思维，并引发技术和经济变革。
最后，他提到了脑机接口技术，并认为虽然它是一种接口，但可能会存在其他更好的压缩信息的方法，找到比语言更高效的沟通方式。他还给AI研究人员们提供了一些建议：保持定期写作的习惯、对流行趋势要保持中立、选择你认为既重要又可能有研究成果的问题去研究。
这篇文章为我们带来了关于理查德·萨顿对AI研究方向的一些见解和思考，希望能够在探索人工智能的道路上给大家带来一些启发。