I don't get it. Where is the spec? What's the aperture size, focal length of this beast? What about abberations? How does it work in extreme conditions that a normal lens can't take a good image?
So many questions, but the author decided to use it to take portraits in day light ...
I guess different people have different definition for "significant circumstantial evidence". We have two hypotheses: 1. virus come from seafood market, 2. virus come from lab. Let's see the likelihood for each with each piece of data.
Data 1: Index case. There is no evidence that the "index case" is ever found. The earliest patient backtracked to Nov 2019 [1], but there is no report that patient is related to the lab / market.
Data 2: close lab location to seafood market. This alone does not favor one hypothesis to other, since distance is communicative. The virus can start from either location to the other.
Data 3: Bat species. Not sure what aspect of the bat species is related, but one thing to note is the virus sample found in bat is only 96% similar to the virus sample. At 30k genome and an mutation rate of 1bp per two weeks, these samples have at least 20 years worth of evolution time. Unlikely to relate to either lab / seafood market. Some researchers believe there might be intermediate hosts, but there doesn't seem to be evidence what that intermediate host might be.
Data 4. Earlier lab accident. I think the lab accident data is actually not favoring the lab hypothesis. If you think in a Bayesian way, earlier lab leaks are quickly identified and controlled. Given that this time it is not, your belief for lab leak should decrease a bit. Anyway, for the lab data what's important is the likelihood of (the lab, without genetic engineering, get a hold of this virus, and leaked) vs. (an intermediate host). Having worked in Biohazard level 2/3 labs, I think a leak from level 4 lab will be more unlikely compared to intermediate host, but we don't have any good estimate of the two likelihood yet.
So I think people can have strong beliefs about where the virus come from (since everyone's prior is different), but from the data the likelihood really doesn't strongly favor any of the two hypotheses.
> but from the data the likelihood really doesn't strongly favor any of the two hypotheses.
I mostly agree with this. I personally think the totality of the evidence favors the lab hypothesis, for a variety of reasons (note the the other lab conducting and publishing about live bat research in Wuhan - the Wuhan Centers for Disease Control - was BL2, I agree it's unlikely to be the BL4 Institute of Virology).
But I also agree there's also a decent probability it was the market. I'd say 3:1, if I had to put a number to it.
However, folks on the other side seem to want to categorically rule out an accidental lab release, seemingly for some social or political reason. That seems very wrong-headed.
Edit: btw, regarding your point 4, a Bayesian approach would seem to provide more support for the accidental lab release, given the posterior. That's actually the approach I'm using here, I even meant to mention it in one of my comments. That said, I hadn't thought of your interpretation. I'll think about that, but I think that what may be different here is that one of the labs in question here isn't under national control, unlike the previous ones that have had accidental releases. And we all know the problems provincial governments have had with corruption and coverup. But I'll think about what you said.
> Given that this time it is not, your belief for lab leak should decrease a bit.
China's ambitions and role in the World are very different in 2020 than they were in 2004. If China manages to contain this pandemic better than the US and Europe, it'll come out relatively stronger.
We do not have hard evidence about the lab leak hypothesis and we must have it before drawing any conclusions. The timing of the economic repercussions on the US/EU seems to come at a strategically perfect time for China to gain politically. This would be a much better alternative to a way and much easier to cover up. All relevant evidence may have been wiped away, for all we know.
All I am saying is that there are definitely reasons to have different Bayesian priors in 2020 compared to 2004. Evidence is still weak, I agree.
Being someone who is still struggling with these concepts, I like this tutorial because at least it doesn't use the common list/maybe/state to illustrate the concepts. Somehow I feel these concepts are so abstract - unless one is well versed in category theory, maybe only a data approach can prevent people from overfitting these concepts to specific examples.
I would really hope to see a tutorial that have a diverse set of examples and just fmap each example with a light explanation of say, what is a monad in this code and what is not, and because it's a monad we can do this.
Essentially the tutorial can just train a classifier in one's head, and with a nice set of examples maybe the brain can learn a general representation of concepts for the classifier ...
In this series functional concepts are very gently introduced. I feel like it really appreciates the beginners starting point and assumes very little. Is this close to what you want?
I am still pretty new learning haskell, and find this very helpful. Not necessarily agree whether monadic solution is prefered, but I like how it shows how general monad is: it's hard to grasp the generalization from only a couple of simple examples normally found in any haskell book. I hope there are more articles like this showing how these abstract concepts can be applied in somewhat unexpected places.
Just to clarify, these MRI images looks at white matter, which are mostly bundles of long-range myelinated axons. This is more likely a map of major highways in the brain (Or really the map of bundles of major highways in the brain, as MRI only has millimeter resolution).
The connectome at the cellular level is massive. I think part of the mouse visual cortical connectome has been mapped out, and the data is on the order of tens of terabytes.
The article indirectly says it's using special MRI machines with much better than millimeter resolution. So I think your first analogy is likely better than your second.
The mouse genome is only 160 megabytes, and contains the instructions for building the brain as well as building everything else, so the "secret sauce" of how to make an intelligent brain should not be extremely large, once you figure out how to do it. :) A lot of the actual connections must be either random, or encoding things the mouse learnt while growing up.
There are four bases, so one base encodes two bits of information. Eight bits are one byte, so four bases are one byte. 2500 megabases = 625 megabytes. So yeah, Parent was off by a factor of 5-6 :) . But still, that fits on one CD.
Except that currently genomics requires even more information to be encoded - such as quality scores, allele frequencies, phase information, ... - so, depending on the format, this estimate is off by either one or two orders of magnitude still.
Take a bunch of source code. Compile it, obfuscate it, compress it, and encrypt it with AES such that the result is a 160MB blob. Now see how long it takes to figure out what it does, given a computer that costs a lot of money just to load your program and a long time to give you a result. The upper bound on the complexity of DNA as it relates to the complete expression of an organism phenotype is insanely high.
Most developers think of the genome like a big load of source code, and if only we could work out where the if and for statements were we could read it. This is an extremely naive and overconfident point of view; the analogy between source code and genomes is very poor. The genome is coding for proteins (by way of RNA). Those proteins are subject to all of physics (think: electrostatics, hydrophobics, ....), whereas your code is an abstract entity designed to run on a rather simple analogue of a Turing machine. The complexity of life is much harder I am afraid. Though that never seems to stop developers assuming that they can create a crude analogy which explains it. Also, the size is totally wrong; see previous comment.
It's true that genes and proteins is nothing like code, but in the context of understanding the brain, I think that should be cause for optimism, because it means that nature has its hands tied behind its back. The genes can't just contain a description of how the brain should be wired together, because the description also has to be "self-executing"; the entire object must robustly self-assemble just from proteins physically interacting. So although 700 megabytes of mouse genes could potentially contain a lot of stuff, it might be possible to do the same thing much more simply if we can program a digital computer instead.
Like, the connectome for C. elegans has been mapped out; it's can be written down as a 2 megabyte ascii text file. Just the connectivity is not enough to actually reproduce the behavior of the worm, you would also need data about the weight of each connection, but it's still a lot less data than the worm genome (about 25 megabytes---I hope I got the number right this time!). The worm genes also need to contain a lot of additional stuff to build functioning cells internals, etc, stuff which hopefully is irrelevant to the actual cognition.
> whereas your code is an abstract entity designed to run on a rather simple analogue of a Turing machine.
I cannot adequately put the insane laugh required as response to that into text form. So I will only write this and be just as right: going by physics the brain of a mouse can be adequately approximated by a perfect sphere.
The definition of a turing machine is mathematically perfect. No threading, no IO, no error correction, no errors, no asynchronous events, no processes fighting over shared resources, no resources that might or might not disappear at the blink of an eye, in short no nothing. In that it is equivalent to a spherical brain, any complexity relevant to the problem at hand removed.
You're making things far too complex, and confusing the issue, and yourself, as a result. Let's turn to the first sentence from Wikipedia:
"A Turing machine ... manipulates symbols on a strip of tape according to a table of rules"
Whichever programming language you are fond of ultimately reduces to this mode of computation. However, with DNA, RNA, and Proteins, that is not the case. The way that we compute is simplistic compared with the way that biology computes. Thus: the crude analogy in fact hinders understanding, and should be discarded.
the "learning" can come from things like the compounds in various food and so on! "learning" is, in a generalized sense, any non-genetically-bootstrapped environment->body information transference...
I think it really depends on the task. We are just hijacking the brain machinery to do jobs it is not evolved to deal with. If we can find some highly evolved / optimized brain function which reflects the structure of the new job, the brain can process it much more efficiently.
Text might be good for programming because, most of the programming are sequential, text is sequential, and text processing is highly optimized in the brain because language. But in fact, we also use a bit of visual programming (indentation, paragraphs) in text to reflect part of the program structure that are not sequential.
If the program is completely non-sequential, visual tools which reflects the structure of the program are going to be much better than text. For example, if you are designing a electronic circuit, you draw a circuit diagram. Describing a electronic circuit purely in text is not going to be very helpful.
I have thought about the same thing when playing with deep learning at home, until I put electricity bills into the equation. A beast like what the author has built, with full power, will cost $0.1/hour in electricity bill alone at where I live ... which almost equal to the cost of a p2.xlarge spot instance on Amazon.
So if someone else is paying for the electricity, build your own gig, otherwise, Amazon is pretty hard to beat.
The spot pricing is pretty crappy to compare to since you only get to use spare capacity. If you have any project where availability matters then you should compare to the on demand price of $0.9/hour.
If availability doesn't matter, it means you wouldn't run the box full time anyway, in which case your power calculation is too high.
I spent almost equal amount of my life in China / North America now, it's interesting to see and compare what one side think about the other.
In north America, what I get is China is controlled by dictators. Life sucks over there because there is no freedom, a lot of corruption, no respect to rules, and is generally a chaotic mess. While in north America, democracy and freedom is wonderful, and everyone has the power to make changes of society, and in the end the government has to care about people to stay in power.
In China, what I get is the US (Canada is usually less discussed) is controlled by a corrupt and impotent government. Life sucks over there because the majority of the people, who actually has the power, believes the doublespeak of the government and are easily manipulated by the wealth. While in China, people make jokes about the official media, and if the government decides to do something people don't like, people simply ignore it and with enough people (there is always enough people in China) doing this, the government has to care about people's opinion to do anything.
I'm a little worried about the situation of both sides now. They seem to go towards what the opposite believes: on one hand in China, the government starts to get better at propaganda and manipulating public opinion; on the other hand in the States ... there is enough happening in the past year.
Most of the Chinese people I know in the US are well educated programmers. They are not true believers in the Chinese Govt supremacy and wonderfulness. But then probably most people are not that well educated, just like the us. China is a large country and will have many diverse opinions, just like the US. I don't meet that many people in the us who are unthinking believers in the supremacy of our country or believe we are special and god loves us. But if you look at a trump rally, or listen to members of congress speak, its like they are from a different planet.
One question to ask is then: what method is useful to figure out the function of an unknown microprocessor? The engineers sometimes reverse engineer chips with electron microscope, with a perfect understanding of electronic principle, a rudimentary idea of how the microprocessor works, what the high-level function of the chip is, etc. And even equipped with this knowledge cracking a chip still takes a lot of effort, and may not succeed at all.
The tools of a neuroscientist would be laughable in comparison, yet neuroscientists have cracked the auditory code (and you can have an artificial cochlea now), the lower visual system is close to be cracked, we have a pretty good understanding of hippocampus and how space is represented in the brain, and many other accomplishment. These are all results from the laughable tools we have for investigating brain.
Having a full connectome of the brain, and running the brain in a simulator will be a huge step of understanding it. We don't know how to use the data now simply because we don't have the data yet, and therefore no effort is putting on interpreting the data. However the usefulness can be glimpsed from neural structures where the connections are clear: the peripheral system, spinal cord, midbrain, and lower sensory and motor regions in the brain. We understand them far better than regions we don't even know where it connects: claustrum comes into mind.
A simulator of the brain, I imagine, will be similar to the human genome project: nobody will understand the whole thing quickly, but it will hugely prop neuroscience forward, sometimes in ways we cannot imagine now.
So many questions, but the author decided to use it to take portraits in day light ...