TomorrowX conference: Roboflow made an interesting observation about the differences between language and vision processing.
There’s a lot of training commonality with language, whereas vision has a huge long tail problem. Every distinct AI vision application requires unique training on that vision dataset. There is no base LLM that knows all of vision, nor will there likely ever be since such a model would be ginormous.
Add in that vision tasks need speed and low latency meaning on device or at least on prem models meaning smaller models and you’ve got the reason for purpose built, specific application trained models.
Humans get around this by learning on the go. Ie. Continuous learning. Yet another reason why lack of continuous learning is an impediment to AI. AI will still be used and is useful, it’ll just not be as efficient from a training perspective until we have continuous learning.
顯示更多