Probability Micro-chip
5 views
Skip to first unread message
Gus Lott
Aug 18, 2010, 9:08:52 AM8/18/10
to The ALISA Community
http://www.technologyreview.com/computing/26055/
Interesting article on a new processor architecture made of "Bayesian
NAND gates" for direct probabilistic calculations in hardware. Could
be the first part of Mada's cortex?
Interesting article on a new processor architecture made of "Bayesian
NAND gates" for direct probabilistic calculations in hardware. Could
be the first part of Mada's cortex?
* Glenn Becker
Sep 11, 2010, 10:21:18 PM9/11/10
to The ALISA Community
Interesting short article, "Brains and Bytes", in the Sept 2010
Communications of the ACM - http://cacm.acm.org/news/98029-brains-and-bytes/abstract
The article itself is very high-level, but it ends with the Blue
Gene / Blue Brain debate.
The Blue Gene project has been simulating 10**9 neurons with 10**13
synaptic connections, but the neurons are simplified models.
The Blue Brain project, on the other hand, simulates less than 10**5
neurons, but uses a far more complex and accurate neuron model.
Is the goal to understand how the brain works or to build machine
cognition?
The Blue Brain approach seems to be correct if the goal is to
ultimately understand the details of how the biological brain works.
But if the goal is to build machine cognition? Does it matter if the
neurons are based on a biological model, or only that they are
statistical and adaptive?
When building flying machines we realized that the shape of the wing
in biology and machine was important but the flapping wasn't
necessary.
What characteristics of the biological neuron are really important?
.
Communications of the ACM - http://cacm.acm.org/news/98029-brains-and-bytes/abstract
The article itself is very high-level, but it ends with the Blue
Gene / Blue Brain debate.
The Blue Gene project has been simulating 10**9 neurons with 10**13
synaptic connections, but the neurons are simplified models.
The Blue Brain project, on the other hand, simulates less than 10**5
neurons, but uses a far more complex and accurate neuron model.
Is the goal to understand how the brain works or to build machine
cognition?
The Blue Brain approach seems to be correct if the goal is to
ultimately understand the details of how the biological brain works.
But if the goal is to build machine cognition? Does it matter if the
neurons are based on a biological model, or only that they are
statistical and adaptive?
When building flying machines we realized that the shape of the wing
in biology and machine was important but the flapping wasn't
necessary.
What characteristics of the biological neuron are really important?
.
Message has been deleted
Message has been deleted
Message has been deleted
* Peter Bock
Sep 12, 2010, 4:53:04 PM9/12/10
to The ALISA Community
Kudos to Lyric Semiconductor for a truly forward-looking idea, but
poor marks for the editors at MIT's Technology Review and author
(journalist?) Tom Simonite, whose definition of the logic of a NAND
gate is NOT conventional: "a conventional NAND gate outputs a 1 if
neither of its inputs match." Nope, that's a (NOT XOR) gate. I think
we would all agree that a conventional NAND gate outputs a 1 iff
either of its inputs is 0.
Having clarified that, now consider the description of the new
probabilistic NAND gate: "...the output of a Bayesian NAND gate
represents the odds that the two input probabilities match..."? Recall
that in inferential statistics, all you may do is compute the
confidence that the NULL hypothesis (H0) may be REJECTED, or (if you
dare) SPECULATE about the acceptance of the alternative (remember
those dastardly ubiquitous Type II errors?). In general, to be on the
safe side, it is generally safer to "reserve judgement."
With that correction, how do we assert H0 so that a sufficient
statistical confidence that H0 may be rejected leads to a reasonable
conclusion that "the output of a Bayesian NAND gate represents the
odds that the two input probabilities match"? I cannot quite connect
the dots here.
Finally, what on earth does this have to do with Bayes?
Am I missing something here?
poor marks for the editors at MIT's Technology Review and author
(journalist?) Tom Simonite, whose definition of the logic of a NAND
gate is NOT conventional: "a conventional NAND gate outputs a 1 if
neither of its inputs match." Nope, that's a (NOT XOR) gate. I think
we would all agree that a conventional NAND gate outputs a 1 iff
either of its inputs is 0.
Having clarified that, now consider the description of the new
probabilistic NAND gate: "...the output of a Bayesian NAND gate
represents the odds that the two input probabilities match..."? Recall
that in inferential statistics, all you may do is compute the
confidence that the NULL hypothesis (H0) may be REJECTED, or (if you
dare) SPECULATE about the acceptance of the alternative (remember
those dastardly ubiquitous Type II errors?). In general, to be on the
safe side, it is generally safer to "reserve judgement."
With that correction, how do we assert H0 so that a sufficient
statistical confidence that H0 may be rejected leads to a reasonable
conclusion that "the output of a Bayesian NAND gate represents the
odds that the two input probabilities match"? I cannot quite connect
the dots here.
Finally, what on earth does this have to do with Bayes?
Am I missing something here?
* Peter Bock
Sep 12, 2010, 5:40:48 PM9/12/10
to The ALISA Community
Apologies for multiple transmissions of my response to the cntribution
from Gus. That old send button just got away from me..... Peter :-)
from Gus. That old send button just got away from me..... Peter :-)
Reply all
Reply to author
Forward
0 new messages