LLMs are opaque. It’s hard to tell what an LLM “knows”. This is historically pretty strange. Good old fashioned AI systems like Cyc or Watson have a huge explicit database of facts and rules, which can be easily queried by the programmer. If you want to know what Watson knows, you can literally just look. But this transparency comes at a price: stupidity. Cyc and Watson know only the facts and rules that have been explicitly programmed into them; they cannot understand freeform text. In contrast, your favourite LLM has read the internet. While it’s difficult to tell what specifically it knows, clearly it knows a whole lot.
But! What if we could have the best of both worlds? People program Cyc and Watson by typing text into a computer. LLMs can type much faster than people can. What if we just asked an LLM to program its knowledge a good old fashioned interpretable system?
In this week’s talk, Blaine Rogers will review the symbolic distillation literature. He’ll talk about the problems researchers encounter using large language models directly as knowledge bases, and survey attempts to extract their knowledge in a queryable symbolic form. Is the future of AI hidden in its deep past? Join us on Wednesday to find out.
The common practice for training commonsense models has gone from–human–tocorpus–to–machine: humans author commonsense knowledge graphs in order to train commonsense models. In this work, we investigate an alternative, from–machine–to–corpusto–machine: general language models author these commonsense knowledge graphs to train commonsense models.
Our study leads to a new framework, Symbolic Knowledge Distillation. As with prior art in Knowledge Distillation (Hinton et al., 2015), our approach uses larger models to teach smaller models. A key difference is that we distill knowledge symbolically–as text–in addition to the resulting neural model. We distill only one aspect–the commonsense of a general language model teacher, allowing the student to be a different type of model, a commonsense model. Altogether, we show that careful prompt engineering and a separately trained critic model allow us to selectively distill highquality causal commonsense from GPT-3, a general language model.
Empirical results demonstrate that, for the first time, a human-authored commonsense knowledge graph is surpassed by our automatically distilled variant in all three criteria: quantity, quality, and diversity. In addition, it results in a neural commonsense model that surpasses the teacher model’s commonsense capabilities despite its 100x smaller size. We apply this to the ATOMIC resource, and will share our new symbolic knowledge graph and commonsense models.