I'm not sure how that contradicts my statement, but even focusing on cores, I doubt market demand would match this strategy. In the past many a CPU was sold that had locked potential. With AMD you could even unlock it. https://www.makeuseof.com/answ...
Haven't they done this with CPUs. They could make a smaller chip and get better yields, but it would be more expensive. In principle, it is the same with cars. They must have accountants who run the numbers to maximize something that involves profits as a component.
But if you have their tax records and bank statements for the last ten years, the skin colour becomes irrelevant!
So how much information is enough? It might be more than people realize. A machine learning algorithm is based on correlation, so it will probably give the "bad" information some influence. And what happens for a person that doesn't have many records. I guess it's tough luck since any priors start to have a bigger influence.
Same for almost anything. Skin colour rarely matters, and given enough
more direct data on factors that do matter, skin colour will have no
predictive value, so the algorithm will ignore it.
Even if it matters, it might not be fair. Let's say that a group of
the population is on average more poor. Given that a person's
information is somewhat noisy, a good machine learning algorithm that
can determine that a person is a member of a poor group will give that
person a bias towards poor. In other words, that person might have
identical financial records as someone from another group and receive
a different outcome for say a loan approval. This is a rational
(improves accuracy) somewhat Bayesian decision by the algorithm, but
most would say it is unfair.
The report tells us it was six minutes from takeoff
to crash. The manual trim is certainly "strong enough" to deal
with the trim system -- that's what it does.
All they needed to do was make a left turn for Laguardia like they're
going to pick up the milk.
More dollars are casing the same amount of goods and EVERYBODY pays
more for stuff. The problem here is that although the minimum wage
workers do see a pay increase dollar wise, they eventually see a cost
of living increase and fall back to their existing standard of living.
It still works if EVERYBODY pays more for stuff. If people who earn
more than $15/hr also pay some of that price increase then the workers
still come out ahead. It's one pie and this just gives to poor a
bigger slice. That is unless the owners use this as an excuse to
increase profits and do some mild collusion. Unfortunately, given the
power of large corporations, that is a more likely outcome.
Workers are thrilled to see that minimum wage go from say $13 to
$15/hr, right up until they take home their new paycheck and discover
the cost of dinner out just went from $13 to $15. Imagine that. And
that's how inflation works.
Fortunately Trump eats fast food. In other words, if people who earn
more than $15/hr also pay some of that price increase then the workers
still come out ahead. That is unless the owners use this as an excuse
to increase profits and do some mild collusion. Unfortunately, given
the power of large corporations, that is a more likely outcome.
Hypothetically, if 20% of the deaths are from vaccinated kids and only
1% of the kids are vaccinated then don't take the vaccination because
it is increasing the risk of death from the flu. In terms of fruit,
if 20 apples are rotten and 80 bananas are rotten, you might assume
that bananas are more likely to be rotten. But if I started with 20
apples and 10000 bananas you'd be wrong. The apples are 100%
rotten/dead while the bananas are 0.8% rotten. This is probably based
on Bayes
Rule.
Going back to my first statistic, around 80% of the deaths in healthy
children from the flu each year is in non-vaccinated kids. 4 out of
every 5.
For this to be useful, you need to know what percentage of kids
are vaccinated for the flu. According to the
CDC 57.9% of children get vaccinated, so roughly half. This
means that not getting the vaccine roughly increases the chance of
death by a factor of 4. While this is essentially what you said, the
number could be very different depending on the number of kids getting
vaccinated.
The difference between gambling, insurance, and investing in stocks
I'll play.
Gambling is bad because of negative expectation. Investing is good
because of positive expectation. Insurance is negative expectation but
it can still be good because any reasonable monetary utility function
is going to decrease the value of more money.
Yes, that is exactly what I did when I started my post. Lately,
Slashdot AI articles have been steeped in people implying that there
is something special about organic brains, something that ANNs don't
have yet.
I don't think it's so much that there is something special about
organic as it is that there might be something lacking in ANNs. Real
brains have neurotransmitters and electrical impulses. They change
their physical structure over time. They have lots of innate complex
structure. They have cells besides neurons such as glial cells that
might have cognitive function. We don't know how even simple neural
systems work. (See https://en.wikipedia.org/wiki/...)
In defense of my definition: yes, neurons do pattern matching. It
fires for a certain subset, i.e. a pattern of its input space. Take
the original claim I was responding to and apply it to the simplest
insect brain you can think of. Does the system as a whole really do
much more than pattern matching?
If the claim is that a neural structure is sufficient, I would just
say that. Saying it is just pattern matching is a loaded statement
since it makes one think of simple iid function induction which
doesn't capture lots of machine learning let alone what a brain can
do. However, while I'd be OK with the claim that neural structure is
sufficient, that's not really a strong statement. I'm sure one could
model a Turing machine with neural structure, so it's really just
saying a computer is sufficient.
Anyway I do agree with your goal to remove any metaphysicism which
does have a history in AI. I also think it's good research to better
understand the success of deep learning in many contexts including
GANs and LSTM. I apologize if I gave the impression I was attacking
you.
That is a good point. To me, what a single neuron or collection of
neurons does is pattern matching, in the sense that the output of a
neuron or collection thereof can be regarded as identifying a certain
(abstracted, meaningful) pattern within the universe.
I assume you mean artificial neuron, because we don't really know what
a neuron does in this context.
Generative Adversarial Networks (GANs) to me are definitely showing
signs of that. Just look at Nvidia's latest stuff:
I'm not up on this research, but it seems GANs are just exploiting the
success of deep learning in a predator/prey relationship to create
some interesting data. It's probably useful for a range of
applications, but I'm not sure it does much for general AI.
The difference between us and apes in intelligence is huge, but in
terms of biological evolution and makeup we're not that far
apart. Finding the differences between the brain of a bonobo and a
human is pretty hard. If there is a 'special sauce' to intelligence,
it must be a fairly small (yet very significant) variation on the
theme of the chimp biological neural network.
We don't really understand any of this so it's hard to make any solid
conclusions. Also chimps are pretty smart and their brain network is
very complex with a lot of innate structure. I assume most people
would agree adding convolution was an important step for image
classification with deep learning. We are probably going to have to
add a lot more structure.
Consider advances in science like this:
These references are consistent with my claims. They are trying to
understand the structure of these neural systems. I suppose you could
claim that anything that has a neural structure is some type of
pattern matching, but I would say that's a pretty non-standard
definition.
Whatever it is, it's more than just matching patterns.
How do you know that?
I guess it depends on how you define pattern matching. A lot of
machine learning is just labeling examples. This is what many people
mean by pattern matching. In many ways this is more scientific theory
formation, which is a skill which takes intelligence to do well. In
other words, find a function to fit the data.
A more sophisticated intelligent system is probably going to need
more structure for both the problem and the solution. It needs to
build in complexity over time not just compress data into a concise
function. It needs to reason and plan to accomplish goals. It needs
to interact with others including adversaries. In many ways adding
these constraints to the problem is what makes it solvable since it adds
assumptions that can be exploited. General mathematical pattern
induction is impossible.
Intelligence: The ability to formulate an effective initial response
to a novel situation.
While this is a definition, I'm sure there many others that are just
as interesting. Once we have a decent understanding of a mechanism
then we can have a proper definition that matches many of the
characteristics of these folk definitions. This is typical of science.
I would imagine that consensus breeds its own form of
self-reinforcement. That doesn't make it wrong - it just makes it
likely that when it is wrong, the evidence will be shocking, and
possibly dismissed or actively fought. I guess that's baked into the
scientific process of peer review to some extent.
I agree. It probably happens more today than in the past. Little
research bubbles are created. Since things are so specialized, when
you need someone to review a paper, you want someone with that
specialization. Once you have a critical mass of researchers, it's
self-sufficient and self-reinforcing.
In order to accept generally accepted scientific theories without any
more than a lay-person's understanding of an over-simplified
explanation (say, the kind offered in an article in the NYTimes
Science Tuesday section), you have to assume some obvious questions
that pop up have been dealt with and dismissed for good reason. So for
example, when I hear the lay-person's explanation of carbon dating
(that the relative abundance of radioactive isotopes in a sample
indicates how long ago that carbon was incorporated from the
atmosphere when the sample was a living organism), I always ask myself
"doesn't that assume that the relative abundance of those isotopes in
the atmosphere is constant - or at least, that scientists have some
way of knowing how that ratio has changed over time?"
Yes, you need to assume the scientists know what they are doing.
Unfortunately, I don't think all science is created equal (or equally
easy.) The softer (really harder) sciences undermine people's opinion
of all science. Think health or economics. (Please ignore
science/engineering distinction.) They need to be studied and
exploited but their results are often suspect...
I guess all this talk about red-shift brought that up, because I've
always had a similar question about red-shift as a measure of
distance. Again, there's a buried assumption there that we know the
rate of expansion of the universe - or at least if it's not constant,
we know the rate of acceleration. Do we? And if we don't, is it just
the accuracy of measurements like carbon dating and red-shift distance
measurements that's thrown into question - not the concepts?
By all means study it if you think it's fun. However, in some
contexts, you will need to appeal the experts. There is too much
knowledge for one person to digest, and to truly be an expert does
require lots of time. I think an important but often ignored question
is how to pick "experts". Some people are clearly doing it wrong.
Cherry-picking scientific opinions favorable to your own is what
science denialism *is*. The denialism in science denialism isn't a
denial of truth; it's a denial of burden of proof. Science isn't about
truth, it's about evidence. It doesn't care what you believe, it cares
how you back up your claims.
I'm not sure I agree that is the definition of scientific denialism,
but I think you have a point. I'm not a fan of debate where the goal
is not to understand a problem but to win. You can often see this is
action where someone rebuts and the adversary quickly concedes the
point. He clearly knew he was wrong, but was just hoping the other
side was not prepared to address the issue.
A fundamentalist biologists who don't believe in evolution, or Earth
Scientists who believe in a Young Earth aren't automatically bad
scientists, as long as they don't make unsubstantiated claims. In fact
more conventional scientists aren't in much of a different position;
every scientist has *some* heterodox positions, otherwise there'd be
no point. Every scientist wants to be the one that shakes things up,
but they know other scientists are watching them. That's why
scientists sound so equivocal; a good scientist knows others are
watching, eager to pounce on any overstep.
They are "bad" scientists if they don't accept fundamental results
in their area of research. I assume most scientists who
are denialists are doing so in areas outside their expertise. Of
course, good scientists might have some controversial ideas or require
abnormal levels of confirmation, but they do not believe things that
contradict the available evidence without some strong arguments.
Same goes for anthropogenic climate change. Believing in global
cooling, steady state climate (even through divine intervention), or
warming mostly driven through natural climate cycles doesn't make you
a science denier.
I guess it's semantics, but those examples sure sounds like scientific
denialism. If you want to do good science, you need an good
scientific argument to support your position. Divine intervention
does not cut it.
Demanding that those views be treated as equally well-established as
AGW does.
I guess this implies some type of political decision. If one wants
them treated equally then one is probably concerned about action based
on that information. I think, to ground this rationally, one should
use utility theory. What's the probability of the outcomes associated
with the actions? I think many of the "deniers" are often just
skeptics. For them an 1% chance that global warming is wrong is
enough to argue the point. However, in this context, 1% is not
terribly relevant for policy. Just look at how often they accept bets
on global warming. Clearly they don't have high confidence.
I use "the left" in a broad term in this case, but you can round it
out from environmentalism shifting to hyper-environmentalism that
humans need to die.
While I think there is some truth to what you say, I do think the
right often resorts to a false equivalence. It might be easy to find
some hyper-environmentalists and trot them out, but it doesn't mean
they own the mind share of the left. It's also easy to find a group
of crazy teachers and students at various universities, but again,
this is only a small fraction of the left. However, if you look at
the surveys, large portions of the right wing believe crazy/bad/wrong
things. For example, InfoWars gets lots of viewers and was even
endorsed by Trump. I'm sure some people just watch for entertainment
value, but they are very profitable selling their snake oil, so there
are probably a lot of true believers.
no, you just don't do machine learning in this application. you use a
clear audit-able rules based system.
There used to be something called expert systems where experts were
questioned to create rules. This failed as the expert systems never
did nearly as well as the experts. Turns out experts can't articulate
how they are experts. As a relatable if somewhat misleading example,
you could ask me for instructions on how to ride a bike, but I could
never give you a set of instructions that would allow you to get on
the bike and just ride.
Thanks for the link. I think power laws are interesting for the scale
invariance and that might be significant for physics. It would be
interesting if this was important for wealth, but it's not the real
problem. Any economic system with strong inequality is probably bad;
however, it seems hard to avoid with a "free" market. A simple random
model with Gaussian returns will give large (exponential tail)
inequality. This is a model with no intelligence, just random
guessing and unsurprisingly some people get lucky and dominate. I'm
sure things get worse when factoring the investment advantages of
being rich. This might also explain the heavy tails. As random
variables get correlated, the central limit theorem no longer applies
and the tails get heavy. Intuitively, I would think an exponential
tail is worse, but I'm not sure...
If people in the long tail don't have enough resources to have at
least an acceptable standard of living, unrest appears, and they will
revolt and coordinate long enough to remove those at the peak; a
position which then will be occupied by a new batch of privileged.
It could be worse; it could be exponential. Also I'm not sure
what you're graphing here. I assume the rich would be on the long
tail.
I'm also curious as to claims that power laws are inevitable. Do
you have a reasonable theory/model to back that up?
Slashdot Mirror
I'm not sure how that contradicts my statement, but even focusing on cores, I doubt market demand would match this strategy. In the past many a CPU was sold that had locked potential. With AMD you could even unlock it. https://www.makeuseof.com/answ...
Haven't they done this with CPUs. They could make a smaller chip and get better yields, but it would be more expensive. In principle, it is the same with cars. They must have accountants who run the numbers to maximize something that involves profits as a component.
So how much information is enough? It might be more than people realize. A machine learning algorithm is based on correlation, so it will probably give the "bad" information some influence. And what happens for a person that doesn't have many records. I guess it's tough luck since any priors start to have a bigger influence.
Even if it matters, it might not be fair. Let's say that a group of the population is on average more poor. Given that a person's information is somewhat noisy, a good machine learning algorithm that can determine that a person is a member of a poor group will give that person a bias towards poor. In other words, that person might have identical financial records as someone from another group and receive a different outcome for say a loan approval. This is a rational (improves accuracy) somewhat Bayesian decision by the algorithm, but most would say it is unfair.
All they needed to do was make a left turn for Laguardia like they're going to pick up the milk.
Not for long. They move.
It still works if EVERYBODY pays more for stuff. If people who earn more than $15/hr also pay some of that price increase then the workers still come out ahead. It's one pie and this just gives to poor a bigger slice. That is unless the owners use this as an excuse to increase profits and do some mild collusion. Unfortunately, given the power of large corporations, that is a more likely outcome.
Fortunately Trump eats fast food. In other words, if people who earn more than $15/hr also pay some of that price increase then the workers still come out ahead. That is unless the owners use this as an excuse to increase profits and do some mild collusion. Unfortunately, given the power of large corporations, that is a more likely outcome.
Hypothetically, if 20% of the deaths are from vaccinated kids and only 1% of the kids are vaccinated then don't take the vaccination because it is increasing the risk of death from the flu. In terms of fruit, if 20 apples are rotten and 80 bananas are rotten, you might assume that bananas are more likely to be rotten. But if I started with 20 apples and 10000 bananas you'd be wrong. The apples are 100% rotten/dead while the bananas are 0.8% rotten. This is probably based on Bayes Rule.
For this to be useful, you need to know what percentage of kids are vaccinated for the flu. According to the CDC 57.9% of children get vaccinated, so roughly half. This means that not getting the vaccine roughly increases the chance of death by a factor of 4. While this is essentially what you said, the number could be very different depending on the number of kids getting vaccinated.
I'll play.
Gambling is bad because of negative expectation. Investing is good because of positive expectation. Insurance is negative expectation but it can still be good because any reasonable monetary utility function is going to decrease the value of more money.
Maybe vitamin D is not the only reason the sun is good for you. https://www.outsideonline.com/...
I don't think it's so much that there is something special about organic as it is that there might be something lacking in ANNs. Real brains have neurotransmitters and electrical impulses. They change their physical structure over time. They have lots of innate complex structure. They have cells besides neurons such as glial cells that might have cognitive function. We don't know how even simple neural systems work. (See https://en.wikipedia.org/wiki/...)
If the claim is that a neural structure is sufficient, I would just say that. Saying it is just pattern matching is a loaded statement since it makes one think of simple iid function induction which doesn't capture lots of machine learning let alone what a brain can do. However, while I'd be OK with the claim that neural structure is sufficient, that's not really a strong statement. I'm sure one could model a Turing machine with neural structure, so it's really just saying a computer is sufficient.
Anyway I do agree with your goal to remove any metaphysicism which does have a history in AI. I also think it's good research to better understand the success of deep learning in many contexts including GANs and LSTM. I apologize if I gave the impression I was attacking you.
I assume you mean artificial neuron, because we don't really know what a neuron does in this context.
I'm not up on this research, but it seems GANs are just exploiting the success of deep learning in a predator/prey relationship to create some interesting data. It's probably useful for a range of applications, but I'm not sure it does much for general AI.
We don't really understand any of this so it's hard to make any solid conclusions. Also chimps are pretty smart and their brain network is very complex with a lot of innate structure. I assume most people would agree adding convolution was an important step for image classification with deep learning. We are probably going to have to add a lot more structure.
These references are consistent with my claims. They are trying to understand the structure of these neural systems. I suppose you could claim that anything that has a neural structure is some type of pattern matching, but I would say that's a pretty non-standard definition.
I guess it depends on how you define pattern matching. A lot of machine learning is just labeling examples. This is what many people mean by pattern matching. In many ways this is more scientific theory formation, which is a skill which takes intelligence to do well. In other words, find a function to fit the data.
A more sophisticated intelligent system is probably going to need more structure for both the problem and the solution. It needs to build in complexity over time not just compress data into a concise function. It needs to reason and plan to accomplish goals. It needs to interact with others including adversaries. In many ways adding these constraints to the problem is what makes it solvable since it adds assumptions that can be exploited. General mathematical pattern induction is impossible.
While this is a definition, I'm sure there many others that are just as interesting. Once we have a decent understanding of a mechanism then we can have a proper definition that matches many of the characteristics of these folk definitions. This is typical of science.
Haters gotta hate.
I agree. It probably happens more today than in the past. Little research bubbles are created. Since things are so specialized, when you need someone to review a paper, you want someone with that specialization. Once you have a critical mass of researchers, it's self-sufficient and self-reinforcing.
Yes, you need to assume the scientists know what they are doing. Unfortunately, I don't think all science is created equal (or equally easy.) The softer (really harder) sciences undermine people's opinion of all science. Think health or economics. (Please ignore science/engineering distinction.) They need to be studied and exploited but their results are often suspect...
By all means study it if you think it's fun. However, in some contexts, you will need to appeal the experts. There is too much knowledge for one person to digest, and to truly be an expert does require lots of time. I think an important but often ignored question is how to pick "experts". Some people are clearly doing it wrong.
I'm not sure I agree that is the definition of scientific denialism, but I think you have a point. I'm not a fan of debate where the goal is not to understand a problem but to win. You can often see this is action where someone rebuts and the adversary quickly concedes the point. He clearly knew he was wrong, but was just hoping the other side was not prepared to address the issue.
They are "bad" scientists if they don't accept fundamental results in their area of research. I assume most scientists who are denialists are doing so in areas outside their expertise. Of course, good scientists might have some controversial ideas or require abnormal levels of confirmation, but they do not believe things that contradict the available evidence without some strong arguments.
I guess it's semantics, but those examples sure sounds like scientific denialism. If you want to do good science, you need an good scientific argument to support your position. Divine intervention does not cut it.
I guess this implies some type of political decision. If one wants them treated equally then one is probably concerned about action based on that information. I think, to ground this rationally, one should use utility theory. What's the probability of the outcomes associated with the actions? I think many of the "deniers" are often just skeptics. For them an 1% chance that global warming is wrong is enough to argue the point. However, in this context, 1% is not terribly relevant for policy. Just look at how often they accept bets on global warming. Clearly they don't have high confidence.
It looks like you forgot to convert from m^3 to liters. Isn't the right number around 31 MJ/L? So about the same as gas, which is not bad.
While I think there is some truth to what you say, I do think the right often resorts to a false equivalence. It might be easy to find some hyper-environmentalists and trot them out, but it doesn't mean they own the mind share of the left. It's also easy to find a group of crazy teachers and students at various universities, but again, this is only a small fraction of the left. However, if you look at the surveys, large portions of the right wing believe crazy/bad/wrong things. For example, InfoWars gets lots of viewers and was even endorsed by Trump. I'm sure some people just watch for entertainment value, but they are very profitable selling their snake oil, so there are probably a lot of true believers.
There used to be something called expert systems where experts were questioned to create rules. This failed as the expert systems never did nearly as well as the experts. Turns out experts can't articulate how they are experts. As a relatable if somewhat misleading example, you could ask me for instructions on how to ride a bike, but I could never give you a set of instructions that would allow you to get on the bike and just ride.
Thanks for the link. I think power laws are interesting for the scale invariance and that might be significant for physics. It would be interesting if this was important for wealth, but it's not the real problem. Any economic system with strong inequality is probably bad; however, it seems hard to avoid with a "free" market. A simple random model with Gaussian returns will give large (exponential tail) inequality. This is a model with no intelligence, just random guessing and unsurprisingly some people get lucky and dominate. I'm sure things get worse when factoring the investment advantages of being rich. This might also explain the heavy tails. As random variables get correlated, the central limit theorem no longer applies and the tails get heavy. Intuitively, I would think an exponential tail is worse, but I'm not sure...
It could be worse; it could be exponential. Also I'm not sure what you're graphing here. I assume the rich would be on the long tail.
I'm also curious as to claims that power laws are inevitable. Do you have a reasonable theory/model to back that up?
While the US does well compared to China that's a pretty low bar. The real competition is Vatican City. That country is a model for the world.