A lossless compression contest on video would result in contestants spending 99.9999% of their efforts on compressing data that the eye and brain throw away, assuming the payoff is the same for both types. Noise is not just white noise, but all the details in the scene that don't increase your reproductive fitness.
Also, you would need about 10-20 years worth to model human level vision, or 100 exabytes at the eye's resolution. Text compression works because you only need 1 GB, and it's mostly signal. On Sat, Mar 21, 2020, 1:24 PM James Bowery <jabow...@gmail.com> wrote: > > > On Sat, Mar 21, 2020 at 12:00 PM Matt Mahoney <mattmahone...@gmail.com> > wrote: > >> Data compression won't solve AGI. It's just a useful tool for evaluating >> language models... >> > > I'd put it more like "Data compression _alone_ won't solve AGI. It's just > the gold standard for model selection. Model selection _alone_ won't solve > AGI." > > >> Compression is not so useful for evaluating images, video, or audio >> models. >> > > That doesn't mean it's not the gold standard for model selection. > "usefulness" aka "utility"gets one into the Sequential Decision Theory > aspect of AIXI. > > >> Lossless algorithms are overwhelmed by incompressible noise. >> > > As long as two models are challenged with _exactly_ the same data, the > small differences make all the difference in terms of model selection. One > man's noise is another man's ciphertext. > > *Artificial General Intelligence List <https://agi.topicbox.com/latest>* > / AGI / see discussions <https://agi.topicbox.com/groups/agi> + > participants <https://agi.topicbox.com/groups/agi/members> + delivery > options <https://agi.topicbox.com/groups/agi/subscription> Permalink > <https://agi.topicbox.com/groups/agi/T2a0cd9d392f9ff94-M7e108fbea78c86fa66282984> > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T2a0cd9d392f9ff94-M9537928c153ea765257d3706 Delivery options: https://agi.topicbox.com/groups/agi/subscription
