I understand the desire to understand what an AGI knows. But that makes you smarter than the AGI. I don't think you want that.
A neural network learner compresses its training data lossily. It is lossy because the training data information content can exceed the neural network's memory capacity (as all learners should). Then it compresses the remainder effectively by storing as prediction errors. Learning simply means making whatever adjustments reduce the error. On Fri, Oct 5, 2018, 10:29 PM Ben Goertzel <b...@goertzel.org> wrote: > Jim, > > If you look at how lossless compression works, e.g. lossless text > compression, it is mostly based on predictive probability models ... > > If you have an opaque predictive model of a body of text, e.g. a deep > NN, then it's hard to manipulate the internals of the model ... > > OTOH if you have a predictive model that is explicitly represented as > (say) a probabilistic logic program, then it's easier to manipulate > the internals of the model... > > So I think actually "operating on compressed versions of data" is > roughly equivalent to "producing highly accurate probabilistic models > that have transparent internal semantics" > > Which is important for AGI for a lot of reasons > > -- Ben > On Sat, Oct 6, 2018 at 5:05 AM Jim Bromer via AGI <agi@agi.topicbox.com> > wrote: > > > > A good goal for a next generation compression system is to allow > > functional transformations to operate on some compressed data without > > needing to decompress it first. (I forgot what this is called but > > there is a Wikipedia entry on something s8milar in cryptography.) > > This is how multiplication works by the way. > > > > If a 'dynamic compression' was preformed in stages using 'components' > > which had certain abstract attributes that could be used in > > computations that were done in multiple passes, then it might be > > possible to postpone a complete analysis or computation until the data > > was presented in a more abstract format (relative to the given > > problem). The goal is to find a way to make each pass effective but > > seriously less complicated. The idea is that the data 'components' > > (the data produced by a previous pass) might have certain abstract > > properties that were general, and subsequent passes might then operate > > on narrower classes. (This is how many algorithms work now that I > > think about it, but they are not described and defined using the > > concept of compression abstractions as a fundamental principle.) > > Jim Bromer > > > -- > Ben Goertzel, PhD > http://goertzel.org > > "The dewdrop world / Is the dewdrop world / And yet, and yet …" -- > Kobayashi Issa ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T55454c75265cabe2-M850081d8e69a884abd15986d Delivery options: https://agi.topicbox.com/groups/agi/subscription
