I didn't understand exactly why the forms of research this allows for is so interesting or important even? He just claims that the high dimensional visualizations are "obviously important and powerful" but doesn't explain why.
I'm always skeptical of researchers who come out with pretty visualizations actually just looking for funding and/or recognition.
It's interesting as it shows that people are using Spark for something other than for the traditional web/enterprise analytics. Most of the summit was focused on BI related use cases so it was really nice to see something different for a change.
It's important as the speed and interactivity of Spark has apparently helped the lab quite a lot in their research efforts. Some of the things shown in the video, particularly the real-time refocusing at the end of the video, would have been a lot more difficult, if not impossible, without something like Spark (or similar)
I just don't see how this particular application is so important. At most, it seems to be demonstration of sparks capacities but again it was geared for high performance cluster computing anyways.
I can understand your skepticism but Jeremy has done some really phenomenal work beyond visualization.
And in this case, he doesn't have to look for funding since he works at Janelia Farm. Depending on the position, they're not even allowed to apply for outside funding.
Right, well I mean this obviously helps whatever institution (in this case Janelia Farm) get funding for research. I watched the lecture again and I'm very positive that this was more of a "show than tell" talk. As far as I can figure, there honestly wasn't anything that interesting in terms of neuroscience research.
Anyways, I wasn't so much skeptical of Jeremy as much as the talk and its intentions. I did a little bit of research on his work but nothing immediately stood out. What's some of the research he's done that you would personally point me to and crucially why? I'll do my part and investigate it.
While I think the presentation was interesting it was fundamentally just more proof that "the brain has patterns". Wish that the talk had more depth in terms of understanding or implementing artificial neural networks. Understanding that the presentation was primarily to endorse Spark to more data mining applications however, the talk seems successful.
>>terms of understanding or implementing artificial neural networks
There is no motivation for a connection between artificial and biological computational networks. A production ANN such as those used by missile interception systems is not requires to match what the neurons in your brain do. In fact, the biologists may find that our brains are not very efficient and often make mistakes. As a practical matter, the behavior of neurons is still up for debate with many models of interaction proposed.
There are often more effective models for machine learning, especially if you have lots of data, such as Random Forest.
>>There is no motivation for a connection between artificial and biological computational networks.
I would disagree, and say that future ANN research has high motivation to become as powerful and complex as BNNs. And I think maybe you get ahead of yourself by saying that our brains are not very efficient when the fastest supercomputer we currently have built is still 100,000x less computationally powerful than a human brain.
"Neural Networks attempt to bring computers a little closer to the brain’s capabilities by imitating aspects of
information in the brain in a highly simplified way. Although neural networks as they are implemented on computers were inspired by the function of biological neurons, many of the designs have become far removed from biological reality."
>There is no motivation for a connection between artificial and biological computational networks.
Not sure what you mean by "connection", but there is a huge motivation for understanding the biological. The neocortex is the single best implementation of intelligence we know.
You should watch a couple of video talks by Jeff Hawkins which are REALLY convincing. His work on sparse distributed representation is, in my opinion, one of the biggest recent break throughs in artificial intelligence that was directly achieved by researching the neocortex.
some of their analyses are performed on raw voxels, but individual neurons for ablation experiments can be identified either manually or using video image processing to identify single neurons. It is a hard problem and a number of people are working on improving it, but for many analyses it may not be strictly necessary as a preprocessing step.
Also I'm in the camp that trying to reverse engineer the brain by studying neural activity is like trying to reverse engineer Angry Birds by studying register activity on an iphone.
Quite. The Blue Brain project did a similar thing by mapping all the neurons in the neocortical column of a cat (or rat?) and visualizing the neural activity and it hasn't yielded anything ground breaking.
Here's Henry Markram's Ted Talk: https://www.youtube.com/watch?v=LS3wMC2BpxU. But honestly, its really just something to impress people who don't have any idea of what kind of implications this would have.
But I do think that obviously understanding neural activity is important but only if its going to help yield understandings of the basic structural units of the neocortex. Trying to understanding the neural activity of millions upon billions is definitely not going to go anywhere.
The approach taken by Markram is quite different from what Jeremy is presenting here. Markram is attempting to simulate large systems, and this work is controversial, largely because many people don't think we understand neural systems well enough to model them with appropriate fidelity at that scale.
The work presented here represents a new take on a more classically accepted approach. Record an animal's behavior as well as the neural activity, and attempt to understand how neural activity controls the observed behaviors. The difference here is a dramatic increase in scale and computational capability by recording the entire brain of a zebrafish at once, and being able to run analyses in a few seconds instead of hours, which has real implications for experimental neuroscientists.
You're right, I made a mistake in implying that the "Markram did a similar thing...".
Specifically I'm criticizing trying to analyze the neural activity of entire brain regions. Neural simulation or actually trying to record neural activity wasn't really my point. I assume both currently have drawbacks weather its "not having accurate enough neural models for simulations" or "not sufficiently or accurately being able to record neural activity".
Yes being able to run analysis in a few seconds instead of hours has real implications for experimental analysis but again the benefits of the technology were never under my criticism or skepticism.
I suppose what I'm specifically alluding to is my criticism of trying to do such large scale analysis of neural activity itself given how little we know about what we should be trying to look for at such a scale.
I say "we" but I should clarify I'm no neuroscientist.
17 comments
[ 3.0 ms ] story [ 48.7 ms ] threadI'm always skeptical of researchers who come out with pretty visualizations actually just looking for funding and/or recognition.
It's important as the speed and interactivity of Spark has apparently helped the lab quite a lot in their research efforts. Some of the things shown in the video, particularly the real-time refocusing at the end of the video, would have been a lot more difficult, if not impossible, without something like Spark (or similar)
Most, if not all, of there projects (https://amplab.cs.berkeley.edu/projects/) are geared at solving similar problems.
I just don't see how this particular application is so important. At most, it seems to be demonstration of sparks capacities but again it was geared for high performance cluster computing anyways.
And in this case, he doesn't have to look for funding since he works at Janelia Farm. Depending on the position, they're not even allowed to apply for outside funding.
Anyways, I wasn't so much skeptical of Jeremy as much as the talk and its intentions. I did a little bit of research on his work but nothing immediately stood out. What's some of the research he's done that you would personally point me to and crucially why? I'll do my part and investigate it.
There is no motivation for a connection between artificial and biological computational networks. A production ANN such as those used by missile interception systems is not requires to match what the neurons in your brain do. In fact, the biologists may find that our brains are not very efficient and often make mistakes. As a practical matter, the behavior of neurons is still up for debate with many models of interaction proposed.
There are often more effective models for machine learning, especially if you have lots of data, such as Random Forest.
I would disagree, and say that future ANN research has high motivation to become as powerful and complex as BNNs. And I think maybe you get ahead of yourself by saying that our brains are not very efficient when the fastest supercomputer we currently have built is still 100,000x less computationally powerful than a human brain.
http://www.wired.com/images_blogs/wiredscience/2013/05/neuro...
"Neural Networks attempt to bring computers a little closer to the brain’s capabilities by imitating aspects of information in the brain in a highly simplified way. Although neural networks as they are implemented on computers were inspired by the function of biological neurons, many of the designs have become far removed from biological reality."
http://scholarsresearchlibrary.com/EJAESR-vol2-iss1/EJAESR-2...
Not sure what you mean by "connection", but there is a huge motivation for understanding the biological. The neocortex is the single best implementation of intelligence we know.
You should watch a couple of video talks by Jeff Hawkins which are REALLY convincing. His work on sparse distributed representation is, in my opinion, one of the biggest recent break throughs in artificial intelligence that was directly achieved by researching the neocortex.
Also I'm in the camp that trying to reverse engineer the brain by studying neural activity is like trying to reverse engineer Angry Birds by studying register activity on an iphone.
Nevertheless, very cool visualizations.
Here's Henry Markram's Ted Talk: https://www.youtube.com/watch?v=LS3wMC2BpxU. But honestly, its really just something to impress people who don't have any idea of what kind of implications this would have.
But I do think that obviously understanding neural activity is important but only if its going to help yield understandings of the basic structural units of the neocortex. Trying to understanding the neural activity of millions upon billions is definitely not going to go anywhere.
The work presented here represents a new take on a more classically accepted approach. Record an animal's behavior as well as the neural activity, and attempt to understand how neural activity controls the observed behaviors. The difference here is a dramatic increase in scale and computational capability by recording the entire brain of a zebrafish at once, and being able to run analyses in a few seconds instead of hours, which has real implications for experimental neuroscientists.
Specifically I'm criticizing trying to analyze the neural activity of entire brain regions. Neural simulation or actually trying to record neural activity wasn't really my point. I assume both currently have drawbacks weather its "not having accurate enough neural models for simulations" or "not sufficiently or accurately being able to record neural activity".
Yes being able to run analysis in a few seconds instead of hours has real implications for experimental analysis but again the benefits of the technology were never under my criticism or skepticism.
I suppose what I'm specifically alluding to is my criticism of trying to do such large scale analysis of neural activity itself given how little we know about what we should be trying to look for at such a scale.
I say "we" but I should clarify I'm no neuroscientist.