François CholletVerified account

@amiconfusediam a beautiful marriage would be Numenta's algorithms actually outperforming benchmarks on somewhat standard tasks.

@Numenta @amiconfusediam "somewhat standard" benchmarks. Can these models do anything sensible on non-synthetic, non-toy data?

@egrefen @amiconfusediam the NAB benchmark has a real-world dataset with 47 real data streams, labeled anomalies. Rewards early detection.

@Numenta @amiconfusediam Cool. Something like MNIST digit labelling or language modelling should be a piece of cake then, no?

@egrefen @Numenta anyhow, it's cool they're doing something different -from an evolutionary standpoint, we don't want everyone working on DL

@fchollet @Numenta They're totally free to do so. Would just be nice to see results on something we can relate other models to now and then.

@egrefen @Numenta Definitely. I see these approaches as a hedge: low success probability, but it's nice to have them around.