> Seeing as how they didn't link to or even cite any of these studies, I think I'll reserve judgment.
Right. No refs combined with an absolutely dumb-downed and/or misinterpreted conclusion to show a whole swath of "stupid" science. As someone entering life in science (the neuro variety), this makes me pretty mad.
I looked into the Psychonomic Bulletin article and it turns out someone I'm collaborating with published in that same issue (Kahana and Howard, "Spacing and lag effects in free recall of pure lists", pp. 159-164). It's not the top psychology journal out there, but it is respectable. The article mentioned in the TFA is:
Loftus, Geoffrey R.; Harley, Erin M. "Why is it easier to identify someone close than far away?". Psychonomic Bulletin & Review, February 2005, vol. 12, no. 1, pp. 43-65(23)
I wonder why the TFA author didn't choose another paper in the same issue to ridicule: "What makes working memory spans so predictive of high-level cognition?". Hmm, maybe because it sounds more intelligent and less stupidly obvious than the Loftus paper, even though it has a silly title-as-question.
The first sentence in the *abstract* of the Loftus paper is:
It is a matter of common sense that a person is easier to recognize when close than when far away.
Ok, the authors know this is common sense. Let's proceed to the second sentence:
First, the human visual system, like many image-processing devices, can be viewed as a spatial filter that passes higher spatial frequencies, expressed in terms of cycles/degree, progressively more poorly.
Ah, so their very first assumption, their starting point for the research reported in the paper, is that perceptual processing sucks for things that are far away (i.e., are smaller in the visual field and thus are characterized by high spatial frequencies). The TFA author ridiculed this as the CONCLUSION of the entire work.
In fact, if you know anything about perceptual cognition, you know that human perception of faces is a very specialized process. So the questions of WHY and HOW perceptual face recognition change with distance may be quite complicated and certainly non-obvious from the start. It's (a priori) a very different question than visual recognition of fruit or plants or car parts or random stimuli.
I don't have access to the full article, but the abstract suggests these findings: demonstrated equivalent performance at equivalent information content (that is, equating low-pass filtering and visual field extension for information) and found evidence for two distinct spatial filters in the visual system (operating at a quantified relative factor) for different perceptual tasks. The discussion relates this to possible models of face perception and real-world applications (eyewitness situations primarily, since they *quantified* where face recognition/perception becomes unreliable).
Now this is the description from TFA:
But surely we can do better than a February study in the journal Psychonomic Bulletin & Review that concluded that it's easier to identify someone close to you than someone more than a football-field-length away. At 450 feet, the scientist concludes, "the human visual system starts to lose small details."
The Wall Street Journal has clearly taken up a strong anti-scientific stance with work like this, and I hope that it doesn't continue. It's abominable. And that was just one of the studies menti^H^H^H^H^Habused.
Besides, even beginning to speak of such things as "units" of consciousness is making many assumptions. I have an "itchy feeling" that the big C arises from a tremendously complex interaction across the many levels of analysis of brain (or "nEUral") structures (from protein phosphorylation to systems topography). The best unit we have to start with is the neuron, and thus neuron theory. They are clearly a computational unit, but nothing suggests an equivalently clear "unit" of consciousness.
There was a great book a couple years ago called "Cradle to Cradle", by William McDonough and Michael Braungart (Slashdot review). It's worth looking into if you're interested in some of the more solid thinking behind "green" design and manufacturing. It's a fun read, and the book is made entirely of recyclable plastic. =)
There aren't very many companies yet embracing the idea of closing off the loop (ie, waste as energy, 100%), but a few are going a long way.
To say that "no one understands how the hippocampus encodes information" is to admit to not even glancing at the research literature. Physiological work dates back to the 70's, revealing mechanisms of hippocampal place cell formation and episodic (declarative) memory formation (e.g., "cognitive map" theories, O'keefe, Nadel, Dostrovsky...). There is definitely some understanding of the Hebbian-like rules (correletional activity strengthens synaptic connections.. or "neurons that fire together wire together") which are critical to the formation of autoassociative memories. Most notably, the modeling and physiology research into "spike-timing dependent plasticity" has been enormous over the last decade at least. This is the idea that the strengthening and weakening of synapses (in cortical and parahippocampal systems) depends on the interval between the firing of the presynaptic and the postsynaptic neuron.
There is also great recent literature on the role of the theta and gamma cycles in hippocampal memory formation (for modeling, see John Lisman at Brandeis, Mike Hasselmo at Boston U.; for the physiology see e.g. Matt Wilson at MIT). The general idea here is that dentate gyrus provides heteroassociative feedback to hippocampus (CA3 region, specifically) on the gamma cycle while the recurrent architecture of CA3 itself provides the autoassociative capabilities necessary for partial-pattern completion (both of which are necessary for declarative sequence learning). Hippocampal place cells (google for more info) are a form of contextual encoding which has been extensively modeled, in general, by e.g. Levy at U.Va (at whose lab I've worked the last few years.)
One of the most important hippocampal functions, though, is its role in memory consolidation, which involves a complex dialog of sorts with neocortex during slow wave sleep. The artificial hippocampus (it's hard to tell if they're replacing the entire thing or just one of the layers) would need to correctly carry out this "teachback" process (which is not understood very well at all). Long story short, even as a relatively contained system, building a silicon replacement hippocampus is not something a budding neuroprostheticist should realistically concern him/herself with.
As kennorman said, the long- and short-term plasticity of the system (i.e., how synapses and neuronal properties change with experience) is far too complex to implement in silicon. I mean, even computational modeling of the system still has a long long way to go. Every researcher has their own model with its own plasticity rules, and as far as most neurobiologists are concerned they're all wrong. Brute-forcing all of the input-output functions of the different cell types is kind of ridiculous for several reasons: 1) most of these i/o functions have been already mapped out by cell physiologists and described by linear-nonlinear models (like double-E's use), 2) the precise architecture (connectivity, topography, etc) of the hippocampal neural networks is "part" of the computation, and 3) hippocampal behavior is dependent on many external factors such as whether or not the brain is sleeping, what hormones happen to flitting around at the moment, etc. Fixed input-output functions will only isolate the behavior of the prosthetic, when it should be a civil member of the society of the brain (otherwise who knows what could go wrong?).
At least though, we actually do know a good deal about hippocampal memory formation, but it's all still just a candle in a dark room.
It's well worth checking out Mike Hasselmo'sarticles, and especially the review article on hippocampal models that he did with Jay McClelland (who I got the chance to meet a few weeks ago, and is the coolest/smartest guy in the field).
Slashdot Mirror
> Seeing as how they didn't link to or even cite any of these studies, I think I'll reserve judgment.
Right. No refs combined with an absolutely dumb-downed and/or misinterpreted conclusion to show a whole swath of "stupid" science. As someone entering life in science (the neuro variety), this makes me pretty mad.
I looked into the Psychonomic Bulletin article and it turns out someone I'm collaborating with published in that same issue (Kahana and Howard, "Spacing and lag effects in free recall of pure lists", pp. 159-164). It's not the top psychology journal out there, but it is respectable. The article mentioned in the TFA is:
I wonder why the TFA author didn't choose another paper in the same issue to ridicule: "What makes working memory spans so predictive of high-level cognition?". Hmm, maybe because it sounds more intelligent and less stupidly obvious than the Loftus paper, even though it has a silly title-as-question.
The first sentence in the *abstract* of the Loftus paper is:
Ok, the authors know this is common sense. Let's proceed to the second sentence:
Ah, so their very first assumption, their starting point for the research reported in the paper, is that perceptual processing sucks for things that are far away (i.e., are smaller in the visual field and thus are characterized by high spatial frequencies). The TFA author ridiculed this as the CONCLUSION of the entire work.
In fact, if you know anything about perceptual cognition, you know that human perception of faces is a very specialized process. So the questions of WHY and HOW perceptual face recognition change with distance may be quite complicated and certainly non-obvious from the start. It's (a priori) a very different question than visual recognition of fruit or plants or car parts or random stimuli.
I don't have access to the full article, but the abstract suggests these findings: demonstrated equivalent performance at equivalent information content (that is, equating low-pass filtering and visual field extension for information) and found evidence for two distinct spatial filters in the visual system (operating at a quantified relative factor) for different perceptual tasks. The discussion relates this to possible models of face perception and real-world applications (eyewitness situations primarily, since they *quantified* where face recognition/perception becomes unreliable).
Now this is the description from TFA:
The Wall Street Journal has clearly taken up a strong anti-scientific stance with work like this, and I hope that it doesn't continue. It's abominable. And that was just one of the studies menti^H^H^H^H^Habused.
"Neuron", dammit. "Neuron"!!
Besides, even beginning to speak of such things as "units" of consciousness is making many assumptions. I have an "itchy feeling" that the big C arises from a tremendously complex interaction across the many levels of analysis of brain (or "nEUral") structures (from protein phosphorylation to systems topography). The best unit we have to start with is the neuron, and thus neuron theory. They are clearly a computational unit, but nothing suggests an equivalently clear "unit" of consciousness.
There was a great book a couple years ago called "Cradle to Cradle", by William McDonough and Michael Braungart (Slashdot review). It's worth looking into if you're interested in some of the more solid thinking behind "green" design and manufacturing. It's a fun read, and the book is made entirely of recyclable plastic. =)
There aren't very many companies yet embracing the idea of closing off the loop (ie, waste as energy, 100%), but a few are going a long way.
To say that "no one understands how the hippocampus encodes information" is to admit to not even glancing at the research literature. Physiological work dates back to the 70's, revealing mechanisms of hippocampal place cell formation and episodic (declarative) memory formation (e.g., "cognitive map" theories, O'keefe, Nadel, Dostrovsky...). There is definitely some understanding of the Hebbian-like rules (correletional activity strengthens synaptic connections.. or "neurons that fire together wire together") which are critical to the formation of autoassociative memories. Most notably, the modeling and physiology research into "spike-timing dependent plasticity" has been enormous over the last decade at least. This is the idea that the strengthening and weakening of synapses (in cortical and parahippocampal systems) depends on the interval between the firing of the presynaptic and the postsynaptic neuron.
There is also great recent literature on the role of the theta and gamma cycles in hippocampal memory formation (for modeling, see John Lisman at Brandeis, Mike Hasselmo at Boston U.; for the physiology see e.g. Matt Wilson at MIT). The general idea here is that dentate gyrus provides heteroassociative feedback to hippocampus (CA3 region, specifically) on the gamma cycle while the recurrent architecture of CA3 itself provides the autoassociative capabilities necessary for partial-pattern completion (both of which are necessary for declarative sequence learning). Hippocampal place cells (google for more info) are a form of contextual encoding which has been extensively modeled, in general, by e.g. Levy at U.Va (at whose lab I've worked the last few years.)
One of the most important hippocampal functions, though, is its role in memory consolidation, which involves a complex dialog of sorts with neocortex during slow wave sleep. The artificial hippocampus (it's hard to tell if they're replacing the entire thing or just one of the layers) would need to correctly carry out this "teachback" process (which is not understood very well at all). Long story short, even as a relatively contained system, building a silicon replacement hippocampus is not something a budding neuroprostheticist should realistically concern him/herself with.
As kennorman said, the long- and short-term plasticity of the system (i.e., how synapses and neuronal properties change with experience) is far too complex to implement in silicon. I mean, even computational modeling of the system still has a long long way to go. Every researcher has their own model with its own plasticity rules, and as far as most neurobiologists are concerned they're all wrong. Brute-forcing all of the input-output functions of the different cell types is kind of ridiculous for several reasons: 1) most of these i/o functions have been already mapped out by cell physiologists and described by linear-nonlinear models (like double-E's use), 2) the precise architecture (connectivity, topography, etc) of the hippocampal neural networks is "part" of the computation, and 3) hippocampal behavior is dependent on many external factors such as whether or not the brain is sleeping, what hormones happen to flitting around at the moment, etc. Fixed input-output functions will only isolate the behavior of the prosthetic, when it should be a civil member of the society of the brain (otherwise who knows what could go wrong?).
At least though, we actually do know a good deal about hippocampal memory formation, but it's all still just a candle in a dark room.
It's well worth checking out Mike Hasselmo's articles, and especially the review article on hippocampal models that he did with Jay McClelland (who I got the chance to meet a few weeks ago, and is the coolest/smartest guy in the field).