Thursday, October 27, 2016
Other ways to count slots in a narrative
You can count all the VARs appearing in a NAR and that is what I am doing. You just climb into the tree and return by adding sub counts. You do the same to count the "used" slots by checking the state of the underlying VARs and adding up the sub counts.
However it might make sense to consider weighting sub narratives equally, even when they contain different numbers of slots. EG if a two part narrative (A,B) has A=(X,Y,Z) and B=(U,V). Then the count of slots for (A,B) should be 1/2*(count for (X,Y,Z)) + 1/2*(count for(U,V)) = 3/2 + 2/2=2.5
The "used" would be weighted in the same way.
Update: Or you could still have the total "amount" of slots equal to 5, but with 1/2 of the 5 devoted to A,B equally, and the other half to X,Y,Z equally. So the slots have individual weights as follows:
A and B each has weight 2.5/2 = 1.25
X, Y, and Z each has weight 2.5/3 = .8333...
However it might make sense to consider weighting sub narratives equally, even when they contain different numbers of slots. EG if a two part narrative (A,B) has A=(X,Y,Z) and B=(U,V). Then the count of slots for (A,B) should be 1/2*(count for (X,Y,Z)) + 1/2*(count for(U,V)) = 3/2 + 2/2=2.5
The "used" would be weighted in the same way.
Update: Or you could still have the total "amount" of slots equal to 5, but with 1/2 of the 5 devoted to A,B equally, and the other half to X,Y,Z equally. So the slots have individual weights as follows:
A and B each has weight 2.5/2 = 1.25
X, Y, and Z each has weight 2.5/3 = .8333...
Tuesday, October 25, 2016
Monday, October 24, 2016
Google's Cloud Analytics
The hype: "Understand the sentiment in a block of text."
My problem is this: blocks of text do not contain a sentiment but a mixture of different sentiments, about different things. No way they would catch the difference between "and" and "but". Talk about "a dull knife"!
Actually: The idea of doing statistical analysis of blocks of text to derive a (non-existent) average sentiment is utter nonsense. It is guaranteed such "tools" will be useless. How long must we toil under an emperor with no clothes? I cannot express the level of distress I have about "researchers" who pass their text data through a black statistical box without understanding the math they are using and without bothering to read the text either. If they read the text, it would quickly become evident that opinions, if subtle, are mixed: consumers "like A" and "dislike B" all at the same time. So trying to turn this into an average opinion is pointless....except to the people who think they are doing research, using off-the-shelf garbage and maybe even getting PhD's. Weirdly enough, companies like Microsoft brag about tools that incorporate "decision tree algorithms". That is 2 strikes against them at the start: using crap math [decision trees aren't even good statistics], and believing in average opinions.
The poor innocent public thinks AI is a done deal and will be more and more in our lives. I strongly suspect the emperor will prevent other contenders for the throne - causing lasting damage, well beyond the short sighted, Bayesian, dreams of the moment.
My problem is this: blocks of text do not contain a sentiment but a mixture of different sentiments, about different things. No way they would catch the difference between "and" and "but". Talk about "a dull knife"!
Actually: The idea of doing statistical analysis of blocks of text to derive a (non-existent) average sentiment is utter nonsense. It is guaranteed such "tools" will be useless. How long must we toil under an emperor with no clothes? I cannot express the level of distress I have about "researchers" who pass their text data through a black statistical box without understanding the math they are using and without bothering to read the text either. If they read the text, it would quickly become evident that opinions, if subtle, are mixed: consumers "like A" and "dislike B" all at the same time. So trying to turn this into an average opinion is pointless....except to the people who think they are doing research, using off-the-shelf garbage and maybe even getting PhD's. Weirdly enough, companies like Microsoft brag about tools that incorporate "decision tree algorithms". That is 2 strikes against them at the start: using crap math [decision trees aren't even good statistics], and believing in average opinions.
The poor innocent public thinks AI is a done deal and will be more and more in our lives. I strongly suspect the emperor will prevent other contenders for the throne - causing lasting damage, well beyond the short sighted, Bayesian, dreams of the moment.
Sunday, October 23, 2016
Coming back from Narwhal's "Darkest Hour"
What I called it here. I have turned the corner and am writing version 3 of the "outer loop" with so many small but difficult lessons learned in the process.
- how scoring could pass from the sub narratives up to the whole, 'linearly'
- how different scoring could be done in the inner loops, separately from return values and from vaulting in the outer loop
- how vaulting does not need to occur when a narrative is complete, as there should be a chance at the next control event
- nor does vaulting need to happen after 'forgetting', for the same reason. [Though there might be a reason to vault after an empty segment of text.]
- that testing every possible segment of text for goodness of fit is WAY too slow, requiring a more relaxed "moving topic" that steps between control indicators in the text. This feels "right"er than version 2.
- how scoring could pass from the sub narratives up to the whole, 'linearly'
- how different scoring could be done in the inner loops, separately from return values and from vaulting in the outer loop
- how vaulting does not need to occur when a narrative is complete, as there should be a chance at the next control event
- nor does vaulting need to happen after 'forgetting', for the same reason. [Though there might be a reason to vault after an empty segment of text.]
- that testing every possible segment of text for goodness of fit is WAY too slow, requiring a more relaxed "moving topic" that steps between control indicators in the text. This feels "right"er than version 2.
Saturday, October 22, 2016
Chatbot to Chatbot Interaction and Community
I am not sure why this isn't an obvious idea but as soon as the software developers realize that bots can interact with other bots then they will be seduced by the artificial life aspect and chatbot development will get at least temporarily side tracked away from the purpose of enhancing my internet experience.
Thursday, October 20, 2016
Two ways chatbots can exchange data
Through their own language interfaces and through standardized data exchange formats. It could be like a secret handshake.
An idea from S Donaldson's "The One Tree"
He repeatedly expresses the idea that when you love something and it is stolen, it is even worse to get it back, broken.
Over on my other blog, I am distressed that readers still are wholly confused about the simple things I have been writing about for years. With whatever sort of spirituality I own, loving the beauty of the woods and all, and with a reasonable sense of the current state of scientific knowledge on this particular subject, I find it repugnant to hear of people lying down in burial mounds and looking up at the stars.
The thing I love is the discovery of these mounds. To show them to people and have them apply ideas which only work for a negligible fraction of the available samples - a population of samples far broader and variable than most people realize - is hard. I am afraid I don't have much patience for it.
Over on my other blog, I am distressed that readers still are wholly confused about the simple things I have been writing about for years. With whatever sort of spirituality I own, loving the beauty of the woods and all, and with a reasonable sense of the current state of scientific knowledge on this particular subject, I find it repugnant to hear of people lying down in burial mounds and looking up at the stars.
The thing I love is the discovery of these mounds. To show them to people and have them apply ideas which only work for a negligible fraction of the available samples - a population of samples far broader and variable than most people realize - is hard. I am afraid I don't have much patience for it.
Wednesday, October 19, 2016
"That" and "It"
I recently put "that" into a list of causality words because the phrase "we were over a bar that was noisy" seems to carry a sense that the bar 'causes' the noise. This sense of causality puts one in a particular narrative framework that is not necessarily related to the old grammatical 'part of speech' definition for the word "that". The 'narrative' reductio is alien to the old grammatical definitions and, hopefully, cuts across them. We'll see, if it turns out I am sorry I put the word into that list.
Update: Perhaps it is simpler to consider "that" to be an attributor like "with", describing a property of the bar. But it is also a property of the experience and its influence on the writer. Still unclear...
UPDATE: Bad idea. It took me a 2 weeks to find out.
Update: Perhaps it is simpler to consider "that" to be an attributor like "with", describing a property of the bar. But it is also a property of the experience and its influence on the writer. Still unclear...
UPDATE: Bad idea. It took me a 2 weeks to find out.
Tuesday, October 18, 2016
Aw AIML!
Going in exactly the wrong direction in some things but, otherwise, probably the authoritative approach to communicating topic context. Gotta get that in Narwhal someday!
I was playing with online demos of ALICE and of CLAUDIO today at Alicebot.org . Here is a fun experiment: what happens if you feed output from ALICE as input to CLAUDIO and vice versa? Do they reach a fixed point? They would have to unless those programs can generate longer and longer texts.
I was playing with online demos of ALICE and of CLAUDIO today at Alicebot.org . Here is a fun experiment: what happens if you feed output from ALICE as input to CLAUDIO and vice versa? Do they reach a fixed point? They would have to unless those programs can generate longer and longer texts.
Monday, October 17, 2016
From an article about Apple hiring AI experts
"Salakhutdinov has published extensively on neural networks, a branch of AI critical for voice and image recognition technology..." [here]
The word "critical" is incorrect and shows the risk of sucking on the bottle of Neural Nets. I am familiar with applying Neural Nets to all of the kinds of image recognition tasks I have encountered in my career*, and I have always felt that neural nets are a disappointment. My whole program is about the necessity for model based recognition. For years, companies like Cognex sold image recognition systems using simple template matching - let alone anything sophisticated. Those work fine, within limits, and use no neural nets.
(*) Including gray scale 2D video in general; video microscopy; character recognition; semiconductor defect recognition; and defect pattern detection; 3D feature recognition.
Update: Also reading about Apple dropping its efforts to develop self-driving cars. Is it that they could not get their neural networks to work, so they are doubling down on that technology? Since AI innovation is not going to come from that direction, I wonder whether there is now no one at Apple with a broad enough technical vision that they are listening to the conventional wisdom about how AI is supposed to be done? In any case it is time to sell your Apple stock and buy IBM.
The word "critical" is incorrect and shows the risk of sucking on the bottle of Neural Nets. I am familiar with applying Neural Nets to all of the kinds of image recognition tasks I have encountered in my career*, and I have always felt that neural nets are a disappointment. My whole program is about the necessity for model based recognition. For years, companies like Cognex sold image recognition systems using simple template matching - let alone anything sophisticated. Those work fine, within limits, and use no neural nets.
(*) Including gray scale 2D video in general; video microscopy; character recognition; semiconductor defect recognition; and defect pattern detection; 3D feature recognition.
Update: Also reading about Apple dropping its efforts to develop self-driving cars. Is it that they could not get their neural networks to work, so they are doubling down on that technology? Since AI innovation is not going to come from that direction, I wonder whether there is now no one at Apple with a broad enough technical vision that they are listening to the conventional wisdom about how AI is supposed to be done? In any case it is time to sell your Apple stock and buy IBM.
Saturday, October 15, 2016
"The Elements of Narrative" published
I have to admit I am a bit pleased. Let's go have a look. Ouch! a cheap looking website. Anyway, here is the email I got:
Congratulations! Your paper has been published at October 2016 issue in IOSR Journal of Engineering.
Congratulations! Your paper has been published at October 2016 issue in IOSR Journal of Engineering.
Your
paper has been published on following link:
Thanks and Welcome!!!
Friday, October 14, 2016
Dreaming about the mathematics
Given the sorts or
semi-mathematical ideas as narrative continuity and invariant transformations of narrative, and given that the entire pattern recognition framework is in the context of a fiber bundle, it is almost guaranteed that
there are Euler numbers and beautiful theorems to find.
Thursday, October 13, 2016
The Goodness of Fit scores I ended up with
During recursive narrative processing, one needs a goodness-of-fit score for a narrative that is linear so the sum of the sub-narrative scores is the same as the score of the whole OR one needs a score that is superimpose-able so the sub-narrative scores can super-impose as the score of the whole. I do both using the U=number of used slots acting as a linear aspect of the score; and the found indices ifound[] as a superimpose-able aspect of the score. At any time you can compare U to the number, N, of slots used. Also you can compare the number of words read, R, to the span of indices (F=last-first+1) where they occur.
If all N slots of a narrative are used then U=N and (U/N)=1. Similarly if every word is read between the first and last indices then (R/F)=1. Thus I propose the formula
But that GOF score is not linear and does not transfer up from the sub-narratives to the whole. So when we come to the need for a goodness of fit score during recursion, the linear/superimpose-able aspects need to used. But how? It only matters when reading the two-part narratives: sequence(a,b) and cause(a,b). What I do is try splitting the text into textA followed by textB and consider U_A as the number of slots used when reading textA with the narrative 'a' and let U_B be the number of slots used when reading textB with the narrative 'b'. Now we seek to maximize
If all N slots of a narrative are used then U=N and (U/N)=1. Similarly if every word is read between the first and last indices then (R/F)=1. Thus I propose the formula
GOF = (U/N)*(R/F)
This is between 0 and 1 and it is equal to 1 if an only if every word is read and every narrative slot is filled. (There are minor adjustments to F, for dull words and known control words. Also the high-level vaulting permits multiple occurrences of the narrative to be counted if they occur repeatedly.)But that GOF score is not linear and does not transfer up from the sub-narratives to the whole. So when we come to the need for a goodness of fit score during recursion, the linear/superimpose-able aspects need to used. But how? It only matters when reading the two-part narratives: sequence(a,b) and cause(a,b). What I do is try splitting the text into textA followed by textB and consider U_A as the number of slots used when reading textA with the narrative 'a' and let U_B be the number of slots used when reading textB with the narrative 'b'. Now we seek to maximize
g = U_A * U_B
over all possible ways of dividing the text into two consecutive pieces. It is tricky because the return value from the reading of this text will be U_A+U_B (using plus! for linearity) where g was maximized. This formula for g favors dividing the text into equal size pieces but the sum does not.
Update: It occurs to me, after explaining that the linear and superimpose-able is preserved in a recursion regardless of what formula you use for g, I can see no reason not to use the full GOF formula for g, as well. I'll have to think about it.
Update: It occurs to me, after explaining that the linear and superimpose-able is preserved in a recursion regardless of what formula you use for g, I can see no reason not to use the full GOF formula for g, as well. I'll have to think about it.
Narwhal is now on GitHub
I believe this is public:
https://github.com/peterwaksman/Narwhal
A bit hard to find using "search" but words like "narhwal" and "tripadvisor" will do it.
https://github.com/peterwaksman/Narwhal
A bit hard to find using "search" but words like "narhwal" and "tripadvisor" will do it.
Tuesday, October 11, 2016
The Internet of Words
I guess I should consider myself the copyright owner of the phrase "Internet of Words". It means an internet powered by language interfaces. I have no intention of enforcing a copyright but hope, rather, that someone does use the phrase.
The Chronicle of Higher Education has me beat but they are referring to something else.
The Chronicle of Higher Education has me beat but they are referring to something else.
Where did all the instant "experts" on chatbots come from?
A few months ago I was writing about the internet of words and was aware of Siri and some efforts underway by Mark Zuckerberg. Now, only a few months later, the use of language interfaces has pole-vaulted into first place in the techworld's discussions of what is cool. What shocks me is to read "Venture Beat" with one article after another about "Chatbots", written by the cool kids who - since they are the authors - seem to think they are experts.
They are experts like grocery store shoppers are experts on canned carrots. I should be pleased that people are beginning to see the internet of words but am a little revolted, as one gets, when personal thoughts get popularized in the mainstream.
I also note with a mixture of fear and skepticism that the highest praise these "experts" have is for chatbots with "... natural language and machine learning features..." and "big data". Fear because deep bullshit is hard to dislodge and skepticism because I am getting closer to launching Narwhal which is a narrow world language processing toolkit, that works with small data and geometry, rather than statistics, as the underlying technology. (I better get that launched asap.)
One last pair of snide comments:
They are experts like grocery store shoppers are experts on canned carrots. I should be pleased that people are beginning to see the internet of words but am a little revolted, as one gets, when personal thoughts get popularized in the mainstream.
I also note with a mixture of fear and skepticism that the highest praise these "experts" have is for chatbots with "... natural language and machine learning features..." and "big data". Fear because deep bullshit is hard to dislodge and skepticism because I am getting closer to launching Narwhal which is a narrow world language processing toolkit, that works with small data and geometry, rather than statistics, as the underlying technology. (I better get that launched asap.)
One last pair of snide comments:
- narrow worlds are exactly what chatbots are for, and so the deep Bayesian statistics approach is doomed.
- it is not about individual chatbots but about how a community of chatbots will work together
And versus But
I always questioned the Symbolic Logic assertion that "and" and "but" had the same meaning. It took years to articulate the basic difference: "and" preserves good/bad value whereas "but" reverses the good/bad value of the previous statement. Similarly "although" reverses the value of the next statment.
But I encountered a more profound difference in the dynamic meaning of "and" versus "but" that is exemplified by how "Bob and Ted and Mary" makes sense where "Bob but Ted but Mary" does not make sense. It is something like this: "but" always signals it was preceded by a complete prior statement and "and" keeps the statement going - allowing it to be preceded by am incomplete prior statement. So the moving read point implementation for "and" does not trigger changes as deep as the "but".
But I encountered a more profound difference in the dynamic meaning of "and" versus "but" that is exemplified by how "Bob and Ted and Mary" makes sense where "Bob but Ted but Mary" does not make sense. It is something like this: "but" always signals it was preceded by a complete prior statement and "and" keeps the statement going - allowing it to be preceded by am incomplete prior statement. So the moving read point implementation for "and" does not trigger changes as deep as the "but".
Sunday, October 9, 2016
Things Squirrels Do
I think I have seen:
- one squirrel dragging another that was wounded out of the road
- a male squirrel fooling with a female squirrel, in missionary position and fondling the tits
- a squirrel placing a leaf across a nest entrance as some kind of signal
Update: namely: compassion, sexual intimacy, and symbolic communication.
- one squirrel dragging another that was wounded out of the road
- a male squirrel fooling with a female squirrel, in missionary position and fondling the tits
- a squirrel placing a leaf across a nest entrance as some kind of signal
Update: namely: compassion, sexual intimacy, and symbolic communication.
Friday, October 7, 2016
Chatbot communities are necessary for them to work together
[Late night reading about "chatbots" - emergent communities of software agents] Narwhal is a design tool for creating conversational interfaces. The whole topic of creating a community of interacting chatbots has gone un-noticed. I presume such a community will work in a way that enhances the total effect, with individual classes able to announce their capabilities to each other and, perhaps, some forms of emergent behavior. The development of environmental controls and of 'learning' features would be interesting challenges.
Thursday, October 6, 2016
First "Goodness Of Fit" measure
I am having a good experience using Microsoft's "Python Tools for Visual Studio". In any case, getting a goodness of fit score ("GOF") for the first time is great. Certainly, I have been aiming for this for months, so getting it to work in one case, is a start. The GOF score is 0.5714...
The text is: "the hotel was near to the border and far from downtown" and the narrative pattern I was looking for was 'room/hotel in proximity to noise sources'. The poor fit is because the word "border" is not in any of the noise source dictionaries. However the word "downtown" is a known noise source.
The text is: "the hotel was near to the border and far from downtown" and the narrative pattern I was looking for was 'room/hotel in proximity to noise sources'. The poor fit is because the word "border" is not in any of the noise source dictionaries. However the word "downtown" is a known noise source.
Tuesday, October 4, 2016
Debt swindle
As far as I can tell, a famous presidential candidate's tax swindle amounts to subtracting a negative and still getting a negative. Where is the bad math? Who took the absolute value of debt?
Monday, October 3, 2016
Glib AI put down
When AI researchers understand the difference between hearing and listening, they will understand what they have been missing. "Deep data" assumes the machine will discover patterns in the data (hearing) but does not understand the need to impose patterns on the data (listening).
Sunday, October 2, 2016
Saturday, October 1, 2016
We were so young... we were so beautiful
Circa 1982 Barbara Jones Waksman:
Circa 1984, Peter Waksman, at USC:
I have a song about Woods Hole:
Circa 1984, Peter Waksman, at USC:
I have a song about Woods Hole:
Do you remember?
Remember when
In the summer
Of two thousand seven
We were so young
We were so beautiful
Let's look back and
Know that we were
Oh, Woods Hole
Is the place of my Dreams
Where my skin's been caressed
By the gentle ocean breeze
And although my friends
Are no longer my friends
I look forward to when
We'll all be friends again
Dumping off "Deep Learning" in Open Source
Here is an example. They use the verb "open source" which I call "dumping off". Either way, Yahoo is not giving up on classifying pornographic images, just their lost investment.
Yahoo open-sources a deep learning model for classifying pornographic images
Subscribe to:
Posts (Atom)