Gary Marcus

Working on a benchmark to test the capacity of contemporary AI to develop anticipate how events unfold over time, given natural language input. If anyone wants to - help make the benchmark or - compete let me know (contact form at garymarcus dot com)https://twitter.com/GaryMarcus/status/1188803198980521986 …

I think a good test would be framed in terms of narrative, which is something we don't have a good formal understanding of yet, but we understand at an intuitive level. Existing tests (e.g. bAbI) are about monitoring factual states of the world. Narrative, on the other hand >

Narrative would represent the changes in state, rather than snapshots. It could best be described as arcs, plots, journeys, and incorporates clear cause-effect relationships. Moreover, it would be centered on agents, including mental models, and how they change over time >

I think such a concept could be very helpful for Cognitive psychology also, which has been nibbling on the edges of this problem for a long time (e.g. schemas, event models), but hasn't had a computational wedge to really dig into it.

I'll go one further and suggest that narrative is an underappreciated organizing principle of human memory function. Which is why the fiction industry is so huge, and we've shared such stories since the dawn of recorded civilization. We just don't understand how it works yet.

It's also relevant to visual scene understanding. I'm not sure how underappreciated it is in memory... lots of people are currently thinking about events/schemas etc. Of course as @GaryMarcus knows this has also been a topic in psycholinguistics for a long time.

That's true, lots thinking about it but I think the mainstay of memory work is still about episodic/semantic. The event cognition stuff is somewhat separate from the classic memory work at the moment.

Maybe I'm just biased by my colleagues here but the event stuff seems to be really taking off in the memory world, especially among people interested in cogneuro and hippocampal/MTL function.