There are two basic approaches to get training data: 1) labeling a visual flow with physics parameters, 2) creating a virtual flow with physics parameters. anyway, by a visual flow with physics parameters, you can train a primitive model of physics pixel firstly. and then apply iternative self-training with human in loop to make the primitive model get better, in which the primitive model generates training data by labeling or creating visual with physics parameters.<\/p>\n\n\n\n
by the way, creating a cyber matrix world of physics pixels is not only good for training but also very cool, isnt it? <\/p>\n\n\n\n
Grok remind me that: synthetic flows (CyberMatrix) are more scalable than real-world labeling.<\/p>\n","protected":false},"excerpt":{"rendered":"
There are two basic approaches to get training data: 1) labeling a visual flow with physics parameters, 2) creating a virtual flow with physics parameters.…<\/p>\n