The claim about sample efficiency sounds a bit strange, since they did not include the state of the art sample efficient algorithms. Like dreamer or tdmpc. Also PPO is known to be not efficient, just compute efficient.

We know that biological neural networks learn faster and orders of magnitude more efficiently than synthetic neural nets. The amount of data/stimulus required to teach a growing human language, motor, social and a variety of other skills, is tiny - when compared to the mass amounts of data required to train SOTA models today.

The question is are there techniques we can adopt from bio neural nets that can enhance the training speed and efficiency of synthetic neural nets?