5 ms·
What people are missing here is that you can setup feedback loops of models to get the "correct" content or whatever you want. For example, If you get too much
by gitfan86 4y ago
What people are missing here is that you can setup feedback loops of models to get the "correct" content or whatever you want.
For example, If you get too much NSFW content put a NSFW detection step on top of GTP ("Is this response offensive") and have that model respond to GTP with "please repeat but with fewer bad words".
For accuracy you can add a Wikipedia feedback loop. "Does this response match information on Wikipedia?" if NO ask it to try again.
Think of these models as transistors or basic logic gates. The real power comes out when you link them together into more advanced logic like a Instruction Pipeline circut.
- jupiterelastica 4y agoLike, a generative model is a source of _some_ information that is refined with gates (classification models) conditional on the generated information? The analogy to transistors and logic gates falls flat a bit when you consider that voltage is a rather simple univariate signal, while generated text is quite complex and multivariate. But I understand that the main point is the composability and filtering.
- gitfan86 4y agoThink of it as information, not voltage. An XOR produces information. A lot of XORS with ANDs make a calculator which opens up an entire vectorspace of mathematical information.
- jupiterelastica 4y agoI try to. One similar thing comes to my mind: generative adversarial networks (GANs). If I'm not mistaken this is along the line of your idea of composing single ML models to bigger information processing units. Do you, by any chance, have links or recommendations for material to read up on architectures that do consider ML models as composable gates?
- gitfan86 4y agoNo materials, it is something I thought up back in March.