[Could a Neuroscientist Understand a Microprocessor? is a riff on the old classic, Can A Biologist Fix a Radio?
Here's one passage that caught my attention.
~Max]
http://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1005268
...What constitutes an understanding of a system? Lazbnick's original paper argued that understanding was achieved when one could "fix" a broken implementation. Understanding of a particular region or part of a system would occur when one could describe so accurately the inputs, the transformation, and the outputs that one brain region could be replaced with an entirely synthetic component. Indeed, some neuroengineers are following this path for sensory [26] and memory [27] systems. Alternatively, we could seek to understand a system at differing, complementary levels of analysis, as David Marr and Tomaso Poggio outlined in 1982 [28]. First, we can ask if we understand what the system does at the computational level: what is the problem it is seeking to solve via computation? We can ask how the system performs this task algorithmically: what processes does it employ to manipulate internal representations? Finally, we can seek to understand how the system implements the above algorithms at a physical level. What are the characteristics of the underlying implementation (in the case of neurons, ion channels, synaptic conductances, neural connectivity, and so on) that give rise to the execution of the algorithm? Ultimately, we want to understand the brain at all these levels.
In this paper, much as in systems neuroscience, we consider the quest to gain an understanding of how circuit elements give rise to computation. Computer architecture studies how small circuit elements, like registers and adders, give rise to a system capable of performing general-purpose computation. When it comes to the processor, we understand this level extremely well, as it is taught to most computer science undergraduates. Knowing what a satisfying answer to "how does a processor compute?" looks like makes it easy to evaluate how much we learn from an experiment or an analysis.
...Lesions studies allow us to study the causal effect of removing a part of the system. We thus chose a number of transistors and asked if they are necessary for each of the behaviors of the processor (Fig 4. In other words, we asked if removed each transistor, if the processor would then still boot the game. Indeed, we found a subset of transistors that makes one of the behaviors (games) impossible. We can thus conclude they are uniquely necessary for the game—perhaps there is a Donkey Kong transistor or a Space Invaders transistor. Even if we can lesion each individual transistor, we do not get much closer to an understanding of how the processor really works.
This finding of course is grossly misleading. The transistors are not specific to any one behavior or game but rather implement simple functions, like full adders. The finding that some of them are important while others are not for a given game is only indirectly indicative of the transistor's role and is unlikely to generalize to other games. Lazebnik [9] made similar observations about this approach in molecular biology, suggesting biologists would obtain a large number of identical radios and shoot them with metal particles at short range, attempting to identify which damaged components gave rise to which broken phenotype.
[Also, don't miss this comment on alpha waves]
In neuroscience there is a rich tradition of analyzing the rhythms in brain regions, the distribution of power across frequencies as a function of the task, and the relation of oscillatory activity across space and time. However, the example of the processor shows that the relation of such measures to underlying function can be extremely complicated. In fact, the authors of this paper would have expected far more peaked frequency distributions for the chip. Moreover, the distribution of frequencies in the brain is often seen as indicative about the underlying biophysics. In our case, there is only one element, the transistor, and not multiple neurotransmitters. And yet, we see a similarly rich distribution of power in the frequency domain. This shows that complex multi-frequency behavior can emerge from the combination of many simple elements. Analyzing the frequency spectra of artifacts thus leads us to be careful about the interpretation of those occurring in the brain. Modeling the processor as a bunch of coupled oscillators, as is common in neuroscience, would make little sense.
[And this part about culture, funding, and goals moving forward is important I think.]
...Culturally, applying these methods to real data, and rewarding those who innovate methodologically, may become more important. We can look at the rise of bioinformatics as an independent field with its own funding streams. Neuroscience needs strong neuroinformatics to make sense of the emerging datasets and known artificial systems can serve as a sanity check and a way of understanding failure modes.
We also want to suggest that it may be an important intermediate step for neuroscience to develop methods that allow understanding a processor. Because they can be simulated in any computer and arbitrarily perturbed, they are a great testbed to ask how useful the methods are that we are using in neuroscience on a daily basis.
--
If I esteem mankind to be in error, shall I bear them down? No. I will lift them up, and in their own way too, if I cannot persuade them my way is better; and I will not seek to compel any man to believe as I do, only by the force of reasoning, for truth will cut its own way.
"Thou shalt love thy wife with all thy heart, and shalt cleave unto her and none else."
http://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1005268
...What constitutes an understanding of a system? Lazbnick's original paper argued that understanding was achieved when one could "fix" a broken implementation. Understanding of a particular region or part of a system would occur when one could describe so accurately the inputs, the transformation, and the outputs that one brain region could be replaced with an entirely synthetic component. Indeed, some neuroengineers are following this path for sensory [26] and memory [27] systems. Alternatively, we could seek to understand a system at differing, complementary levels of analysis, as David Marr and Tomaso Poggio outlined in 1982 [28]. First, we can ask if we understand what the system does at the computational level: what is the problem it is seeking to solve via computation? We can ask how the system performs this task algorithmically: what processes does it employ to manipulate internal representations? Finally, we can seek to understand how the system implements the above algorithms at a physical level. What are the characteristics of the underlying implementation (in the case of neurons, ion channels, synaptic conductances, neural connectivity, and so on) that give rise to the execution of the algorithm? Ultimately, we want to understand the brain at all these levels.
In this paper, much as in systems neuroscience, we consider the quest to gain an understanding of how circuit elements give rise to computation. Computer architecture studies how small circuit elements, like registers and adders, give rise to a system capable of performing general-purpose computation. When it comes to the processor, we understand this level extremely well, as it is taught to most computer science undergraduates. Knowing what a satisfying answer to "how does a processor compute?" looks like makes it easy to evaluate how much we learn from an experiment or an analysis.
...Lesions studies allow us to study the causal effect of removing a part of the system. We thus chose a number of transistors and asked if they are necessary for each of the behaviors of the processor (Fig 4. In other words, we asked if removed each transistor, if the processor would then still boot the game. Indeed, we found a subset of transistors that makes one of the behaviors (games) impossible. We can thus conclude they are uniquely necessary for the game—perhaps there is a Donkey Kong transistor or a Space Invaders transistor. Even if we can lesion each individual transistor, we do not get much closer to an understanding of how the processor really works.
This finding of course is grossly misleading. The transistors are not specific to any one behavior or game but rather implement simple functions, like full adders. The finding that some of them are important while others are not for a given game is only indirectly indicative of the transistor's role and is unlikely to generalize to other games. Lazebnik [9] made similar observations about this approach in molecular biology, suggesting biologists would obtain a large number of identical radios and shoot them with metal particles at short range, attempting to identify which damaged components gave rise to which broken phenotype.
[Also, don't miss this comment on alpha waves]
In neuroscience there is a rich tradition of analyzing the rhythms in brain regions, the distribution of power across frequencies as a function of the task, and the relation of oscillatory activity across space and time. However, the example of the processor shows that the relation of such measures to underlying function can be extremely complicated. In fact, the authors of this paper would have expected far more peaked frequency distributions for the chip. Moreover, the distribution of frequencies in the brain is often seen as indicative about the underlying biophysics. In our case, there is only one element, the transistor, and not multiple neurotransmitters. And yet, we see a similarly rich distribution of power in the frequency domain. This shows that complex multi-frequency behavior can emerge from the combination of many simple elements. Analyzing the frequency spectra of artifacts thus leads us to be careful about the interpretation of those occurring in the brain. Modeling the processor as a bunch of coupled oscillators, as is common in neuroscience, would make little sense.
[And this part about culture, funding, and goals moving forward is important I think.]
...Culturally, applying these methods to real data, and rewarding those who innovate methodologically, may become more important. We can look at the rise of bioinformatics as an independent field with its own funding streams. Neuroscience needs strong neuroinformatics to make sense of the emerging datasets and known artificial systems can serve as a sanity check and a way of understanding failure modes.
We also want to suggest that it may be an important intermediate step for neuroscience to develop methods that allow understanding a processor. Because they can be simulated in any computer and arbitrarily perturbed, they are a great testbed to ask how useful the methods are that we are using in neuroscience on a daily basis.
--
If I esteem mankind to be in error, shall I bear them down? No. I will lift them up, and in their own way too, if I cannot persuade them my way is better; and I will not seek to compel any man to believe as I do, only by the force of reasoning, for truth will cut its own way.
"Thou shalt love thy wife with all thy heart, and shalt cleave unto her and none else."
No comments:
Post a Comment