Also: if you just want a single number out of curiosity: here's what i've got: for just-plain word pairs, I get an entropy of about 18, an MI of -2 (that's a minus sign there) and a sparsity of about 15 (i.e. only 1 in 2^15 entries in the adjacency matrix are non-zero) I don't yet know what these numbers will be like, when I cut out the noise; I have to fix a few bugs first.
For the word-disjunct datasets, I get entropies around 21 and MI of minus 4. and similar sparsities. doubling the size of the dataset has very little effect on these numbers: they might rise by maybe 0.1 or 0.2, but that's it. Fixing bugs that lead to low quality data had the biggest effect: e.g handling punctuation marks correctly helped increase both entropy and MI. p.s. both of those papers make a fundamental error: they state that phi is bounded below by zero. That's clearly and patently false, which would be obvious to the authors if they'd ever applied their formulas to actual data. Boo and big negative marks for making such a basic mistake. --linas On Fri, Jun 16, 2017 at 2:17 PM, Linas Vepstas <[email protected]> wrote: > Matt, Ben, > FYI, > > I've been developing a suite of tools/API's that might be useful for this. > Getting tonini phi is "easy" .. or hard depending on what exactly you > want. So... > > at the core, there's an API to expose a portion of the atomspace as a > matrix. You can think of it as a correlation matrix, covariance matrix, > adjacency matrix of a graph, etc. > > The "adjacency matrix of a graph" is the viewpoint you want for IIT. > > A set of stuff you can do with the matrix, including computing it's > entropy, MI,various RMS deviations. Doing PCA to it. Also, cutting it in > various aways: Currently, I'm using the cuts to filter out noise and junk > data, but you could use the cuts to perform the tonini cuts. > > The "hard part" is that computing entropy for a 30K x 30K matrix can take > hours... so that limits just how many cuts you can explore. > > To use this toolset, you need to write a shim that specifies how to get > the matrix element (i,j) and some related info: the atom type of all the > i's the atom type of all the j's, the atom type of the pair (i,j), and > where to store per-row and per-column data (for subtotals of various kinds) > > I'm actively hacking this now, so its a bit unstable and maybe buggy. Code > is here. > > https://github.com/opencog/atomspace/tree/master/opencog/matrix > > --linas > > On Fri, Jun 16, 2017 at 7:02 AM, Matthew Ikle <[email protected]> wrote: > >> Of course — that is why I called it a pre-experiment. There are some >> subtle mechanics (distribution determination, system partitioning) that we >> need to work out in the actual calculations for phi (or phi*) and the >> approach I outline will allow us to concentrate on figuring these out. Once >> we know how to perform these calculations for this simple pre-experiment, >> then we can more easily add complexity to it by extending and modifying it >> to the AF and/or WA. And of course connecting IIT with strange attractor >> formation would be the “holy grail” of the entire pipeline of experiments, >> >> —matt >> >> > On Jun 15, 2017, at 9:19 PM, Ben Goertzel <[email protected]> wrote: >> > >> > Yes, that would be a meaningful experiment to start with, though >> > obviously not a substitute for an experiment that gauges the >> > integrated information of the dynamics more broadly... >> > >> > On Fri, Jun 16, 2017 at 12:54 AM, Matt Ikle <[email protected]> >> wrote: >> >> Hi Ben, >> >> >> >> I've been thinking a lot about how to actually create experiments for >> >> connecting ECAN with Phi. >> >> >> >> It seems to me that a simple pre-experiment we could run builds upon >> the >> >> insecticide experiment as follows. We could load the insecticide >> experiment >> >> background knowledge and then stimulate the atoms for poison and >> insects >> >> just as before. We can track STI values for these three atoms over >> time and >> >> calculate IIT time series values from the STI time series values. To >> do this >> >> means making only some minor modifications to the practical_phi toolbox >> >> Matlab code to find IIT. >> >> >> >> We may even be able to, though I am more unsure about this, build upon >> this >> >> simple experiment to detect attractor formations within ECAN dynamics >> prior >> >> to looking at the larger ECAN dynamics using PLSI to reduce the space. >> >> >> >> Just trying to create a simple first set of experiments. Thoughts? >> >> >> >> --matt >> >> >> >> Sent from my iPhone >> >> >> >> On May 23, 2017, at 10:18 AM, Ben Goertzel <[email protected]> wrote: >> >> >> >> Yes, agreed... >> >> >> >> On May 24, 2017 00:15, "Matthew Ikle" <[email protected]> wrote: >> >>> >> >>> Yes all of this makes a lot of sense and could be exciting, >> especially the >> >>> connection to Phi. Could be an extremely interesting paper — but a >> lot will >> >>> depend upon parameter tuning as well as creating an appropriate >> experiment. >> >>> >> >>> Before we begin setting up the experiment to look for connections >> between >> >>> Phi and strange attractor structure, though, we obviously must first >> ensure >> >>> basic ECAN implementation as follows: >> >>> >> >>> Step 1: Ensure that ECAN works correctly and fulfills basic design >> >>> criteria; >> >>> Step 2: Run poison experiment and retune parameters. >> >>> >> >>> I feel confident that after Misgana makes the minor changes we >> discussed >> >>> in HK, ECAN should work correctly. Only after we have run through the >> two >> >>> steps above, though. should we proceed with the next (IIT) step and >> perform >> >>> additional parameter tuning. >> >>> >> >>> At some point (probably after all of the above), we should also enable >> >>> HebbianLink updating and run experiments testing the three updating >> >>> equations we have developed and setting the stage for yet another set >> of >> >>> parameter tuning. >> >>> >> >>> —matt >> >>> >> >>>> On May 23, 2017, at 3:52 AM, Ben Goertzel <[email protected]> wrote: >> >>>> >> >>>> Matt, >> >>>> >> >>>> Thinking about how to analyze time-series data from ECAN, I thought >> it >> >>>> might be cool to look for interactions between IIT (Phi) and strange >> >>>> attractor structure in the attentional focus... >> >>>> >> >>>> I found this code which lets us analyze data using Integrated >> >>>> Information Theory (Tononi's Phi) >> >>>> >> >>>> https://figshare.com/articles/phi_toolbox_zip/3203326 >> >>>> >> >>>> This has gotten some acceptance as a "measure of consciousness", so >> if >> >>>> we could show that some ECAN parameters or aspects correlate with >> >>>> "degree of consciousness" as measured by Phi, this would let us >> >>>> publish a wizzy and popular paper.... For instance, what if the >> >>>> system was more conscious (higher Phi) when it connected a sentence >> >>>> with background knowledge, than when it parsed a sentence but was >> >>>> unable to connect it with background knowledge... >> >>>> >> >>>> >> >>>> On the other hand, another interesting thing to do would be to look >> at >> >>>> a delay-embedding of the dynamics... >> >>>> >> >>>> Long ago I used the TISEAN toolkit for nonlinear time series analysis >> >>>> >> >>>> What I am thinking here is: If we are loading in Atoms from a bunch >> of >> >>>> texts, we could run PLSI or similar (latent semantic indexing) on the >> >>>> texts (Eyob could help with that, he's a master of PLSI), to create a >> >>>> dimensional space. At any moment in time, the WordNodes and named >> >>>> ConceptNodes in the AttentionalFocus would then assign the AF a >> >>>> certain point in the dimensional space defined by the PLSI factors. >> >>>> >> >>>> This would turn the AF into a trajectory in n-dimensional space... >> >>>> >> >>>> One could then use some approach to figure out the optimal delay and >> >>>> do a delay-embedding of this trajectory, hopefully revealing the >> >>>> underlying attractor structure... >> >>>> >> >>>> TISEAN seems only to do delay embedding from 1D time series >> >>>> >> >>>> https://www.pks.mpg.de/~tisean/Tisean_3.0.1/index.html >> >>>> >> >>>> but there are papers explaining how to do it from multi-D time series >> >>>> >> >>>> https://arxiv.org/pdf/nlin/0609029.pdf >> >>>> >> >>>> https://arxiv.org/pdf/1409.5974.pdf >> >>>> >> >>>> Showing that the AF contents occupy a certain strange attractor -- >> >>>> maybe shifting which strange attractor over time, or shifting the >> >>>> shape of the strange attractor over time, would be interesting >> >>>> >> >>>> Some association between the Phi (IIT) value and some property of the >> >>>> inferred attractor would also be interesting... >> >>>> >> >>>> -- Ben >> >>>> >> >>>> >> >>>> >> >>>> >> >>>> -- >> >>>> Ben Goertzel, PhD >> >>>> http://goertzel.org >> >>>> >> >>>> "I am God! I am nothing, I'm play, I am freedom, I am life. I am the >> >>>> boundary, I am the peak." -- Alexander Scriabin >> >>> >> >> >> > >> > >> > >> > -- >> > Ben Goertzel, PhD >> > http://goertzel.org >> > >> > "I am God! I am nothing, I'm play, I am freedom, I am life. I am the >> > boundary, I am the peak." -- Alexander Scriabin >> >> -- >> You received this message because you are subscribed to the Google Groups >> "opencog" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to [email protected]. >> To post to this group, send email to [email protected]. >> Visit this group at https://groups.google.com/group/opencog. >> To view this discussion on the web visit https://groups.google.com/d/ms >> gid/opencog/57F4DB20-446A-4565-AA3D-D7C4FFFD3762%40gmail.com. >> For more options, visit https://groups.google.com/d/optout. >> > > -- You received this message because you are subscribed to the Google Groups "opencog" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at https://groups.google.com/group/opencog. To view this discussion on the web visit https://groups.google.com/d/msgid/opencog/CAHrUA35X%2BUMgCjK13YMODAY03L0UcrjL9hjmAmh75gFmCokBBw%40mail.gmail.com. For more options, visit https://groups.google.com/d/optout.
