[8bitsrule]

The buzzword back in the 70s-80s (after AI over-promised in the 1960s) was 'expert systems'. https://en.wikipedia.org/wiki/Expert_system

(To the extent that I have kept up with it) modern AI skips the 'knowledge base' part of ES, in favor of pattern-recognition based on 'training'.

Today's (Indeterministic, trained, n-net) AI has clearly saved a lot of time/effort in creating 'knowledge bases'. I suspect it appeals more to singular fantasies about 'more human than human' intelligence. (Sorry Ray)

Question is: Is today's AI even a magnitude-better than (deterministic) ES insofar as extensibility and verifiability? What if we had spent those decades refining the ES approach instead?

[JimmyRuska]

You should use both. Download wikidata and use the manually curated interconnected data available. You could argue it was compiled by the largest pool of neural networks ever assembled. Wikipedia data was curated content that was hierarchically and painstakingly interconnected by humans, the most advanced NN we know of :).

There's now discussion about how neural networks can succumb to data poisoning / adversarial attacks, because there are no immutable facts. Adding a mostly immutable fact table can help keep things grounded in reason. Most of these engines support complex inference abilities that can lead to unexpected connections.

ES is not really dead. It feels like many rules engines changed their names to "AI Intelligent Agents"-type wording to describe their product. Rete algorithm is similar rule based calculation, is still used to calculate FICO score, which you could say fits into the problems that may be better served by the latest Neural network models. Allegro graph lets you query using prolog and is often used for governance and compliance tools. RDFox is one of the latest inference engines that made major advancements in turning first order logic in datalog into parallel computation.

I'd imagine if you can build a neural network that can successfully interact with a ES knowledge base you could easily make a neural network as good as the one that won in jeopardy

[ilaksh]

I mean, you extend a DL system by giving it examples. Yes, in fact by harvesting example or annotation data you can extend models many magnitudes faster and more comprehensively than through manual analysis. Effectively the analysis is done automatically. At this point, what comes out it is usually poorly factored and entangled and not human-parseable, but that doesn't prevent it from being applied effectively to a range of narrow tasks that expert systems cannot approach.

We do still rely on expert systems for things that we want to be carefully parsed, verified and analyzed by people though. Such as rules that oversee most self-driving cars based on perception handled by neural networks. However, not all self-driving systems lean as heavily on rules at the top level as others.

[jefft255]

They're not even competing with each other. How would you compare an ES to a deep neural network? According to what benchmark? What do you mean by verifiable? Modern success stories with machine learning often deal with problems (NLP, vision, RL) that have little to do with the problems that were being solved by expert systems.

Also, an hypothesis: if (big if) you somehow managed to have an expert system outperform deep learning for vision, I bet that it won't be any more verifiable than a deep neural net is today.

To me the complaint that modern deep learning is unverifiable is a bit dumb, in the sense that any perception algorithm working with low level signals (vision, sound) will not be transparent to a human. 15 years ago, an image classification pipeline looked like: bag of SIFT features + SVM classifier. Try explaining the decision made by that algorithm in an intuitive way!