2 ms·

AI Winter (https://en.wikipedia.org/wiki/AI_winter https://en.wikipedia.org/wiki/AI_winter). 1966: failure of machine translation 1970: abandonment of

AI Winter (https://en.wikipedia.org/wiki/AI_winter https://en.wikipedia.org/wiki/AI_winter). 1966: failure of machine translation 1970: abandonment of connectionism Period of overlapping trends: 1971–75: DARPA's frustration with the Speech Understanding Research program at Carnegie Mellon University 1973: large decrease in AI research in the United Kingdom in response to the Lighthill report 1973–74: DARPA's cutbacks to academic AI research in general 1987: collapse of the LISP machine market 1988: cancellation of new spending on AI by the Strategic Computing Initiative 1993: resistance to new expert systems deployment and maintenance 1990s: end of the Fifth Generation computer project's original goals I got my bachelors in 1990, and took a lot of classes in AI around that time. Have you ever worked with an expert system like Mycin? It is really quite difficult to pull out an expert's knowledge, rules of thumb, and experience-based intuitions. Difficult and expensive. Those that were not tightly focused on a limited domain were also generally not satisfactory, and those that were failed hilariously if any one parameter was outside the system's model. Yes, doctors have a lot of cultural baggage that reduces their effectiveness. But there's a completely different reason why AI has not replaced them. After many, many attempts.

PaulHoule 5y ago
Connectionism is back with a vengeance. It still struggles with text but vision problems like 'detect pedestrian with camera and turn on the persistence-of-vision lightstrip at the right time' are solved. Many expert systems were based on "production rules" and it's a strange story that we have production rules engines that are orders of magnitude more scalable than what we had in the 1980s. Between improved RETE and "look it up in the hashtable" it has been a revolution but production rules have not escaped a few special applications such as banking. Talk to a veteran of "business rules" project and about 50% of the time they will tell you it was a success, the other 50% of the time they made mistakes up front and went into the weeds. Machine learners today repeat the same experiments with the same data sets... That doesn't get you into commercially useful terrain. Cleaning up a data set and defining the problem such that it can be classified accurately is painful in the exact same way extracting rules out of the expert is painful. It's closely related to the concept of "unit test" but it is still a stretch to convince financial messaging experts to publish a set of sample messages for a standard with a high degree of coverage. You can do neat things with text if you can get 1000 to 20,000 labeled samples, but most people give up around 10.

PaulHoule 5y ago

Connectionism is back with a vengeance. It still struggles with text but vision problems like 'detect pedestrian with camera and turn on the persistence-of-vision lightstrip at the right time' are solved. Many expert systems were based on "production rules" and it's a strange story that we have production rules engines that are orders of magnitude more scalable than what we had in the 1980s. Between improved RETE and "look it up in the hashtable" it has been a revolution but production rules have not escaped a few special applications such as banking. Talk to a veteran of "business rules" project and about 50% of the time they will tell you it was a success, the other 50% of the time they made mistakes up front and went into the weeds. Machine learners today repeat the same experiments with the same data sets... That doesn't get you into commercially useful terrain. Cleaning up a data set and defining the problem such that it can be classified accurately is painful in the exact same way extracting rules out of the expert is painful. It's closely related to the concept of "unit test" but it is still a stretch to convince financial messaging experts to publish a set of sample messages for a standard with a high degree of coverage. You can do neat things with text if you can get 1000 to 20,000 labeled samples, but most people give up around 10.