Commentary

Canon T2i/550D, Audio, and Auto Gain Control

4volt — Sat, 01 May 2010 00:02:46 GMT

A while back I posted a glowing review of the video performance of the T2i / 550D camera. I'm coming back now to review the audio quality and frankly there is not much to say.

The external microphone jack works well, and the T2i still uses AGC (Auto Gain Control) that automatically increases the gain (volume) on the microphone when things are quiet. That's fine for impromptu video where you don't necessary know what the noise levels will be, or aren't using an external microphone. The 7D also forces use of AGC but the 5D2 does not since 2.0.4 of the firmware.

For serious or even semi-serious film making, the audio is carefully controlled and microphones are placed to get the best audio quality, and AGC constantly trying to crank up the gain makes things sound very noisy.

What frustrates me most about this is such great manual control of the video on those cameras (now), it makes the audio control and quality seem very poor.

I would love to see Canon add a simple switch to turn of AGC for people that don't want to use it and I've actually started a petition that speaks directly to that.

http://www.petitiononline.com/CanonAGC/petition.html

If you agree, sign the petition, maybe just maybe Canon will listen and add a manual audio control just like they added a manual video control to the 5D Mark II and the 7D after it was released.

Power required to emulate a human brain

4volt — Sat, 06 Feb 2010 07:27:47 GMT

Recently, I posted a article about how much computational power it would take to emulate the human brain, and my calculations were off by a factor of 10. I've corrected my estimates and here are the updated sections:

Emulating the Human Brain

The Blue Brain Project has successfully simulated 1 cordical column on the Magerit super computer (100 TeraFLOPS), and human brain has an estimated 1,000,000 columns.

So with 5.2 ExaFlops, one could emulate 52,000 cordial columns. And that’s 1/20th of enough power to emulate a human brain.

To emulate the entire human brain you would need approximately 100 ExaFLOPs

That's a lot of FLOPS!

I chose the above estimation because many others rely on calculating how many operations per second the brain can processes, not how many computer calculations to emulate the brain. I.E. It takes many more computer computations to emulate a brain computation.

Fore some more notes see: http://www.smartcomputing.com/articles/2002/s1302/39s02/39s02.pdf

Conclusion

Henry Markham in his recent TED talk, thinks that a human brain could be emulated in 10 years.

I agree with his estimate, using Moore's law 10 years seems like an achievable goal.

If you are interested in the subject you may also like Ray Kurzweil's research and books.

Total Processing Power of the Internet

4volt — Mon, 18 Jan 2010 23:38:03 GMT

As a brain exercise I decided to try and estimate the total processing power of all the computers on the entire internet, and see if that is enough processing power to emulate the human brain.

Since this is an estimate, I will try my best to figure it out with public data.

Here was my process:

Total Number of Computers

No one knows exactly how many people (computers) are connected to the internet since a single IP address can be shared with any number of pc’s, but the best estimate I was able to find was here: Internet World Stats

They use population statistics and penetration data to estimate this. This does not include extraneous connected devices like gaming consoles, or other devices.

Total Computers: 1,733,993,741 (1.7 Billion)

Effective Processing Power per Node

Since all the computers are separated by the internet, I chose to use SETI@home as a benchmark for the effective benchmark of how much processing on average a node can handle

According to BOINC stats, Seti has 186,250 active computers (more inactive ones), and the peak processing speed is 704.507 TeraFLOPS (tF). (As of 01/17/2010)

That’s 3.782 GigaFLOPS (gF) per node.

Also, on average computers running SETI are probably faster then the average computer online, especially considering poorer countries. So lets round down the average to 3gF.

3 GigaFLOPS is much lower then the peak processing power of a modern computer, a (Intel Core i7 965 XE), can process about 70, while some of the newest GPU cards can process upwards of 1,000 GigaFLOPS.

This leaves me to believe that tele-processing is very inefficient compared to traditional supercomputers, though it is not without advantage.

Total Processing Power

Now, assume all 1.7B computers were all running this hypothetical peer to peer processing application, and bandwidth was not a bottle neck.

1,733,993,741 * 3 = (5,201,981.223 TeraFLOPS, or 5.2 ExaFLOPS)

In case your wondering:

1 Peta = 1,000 Tera

1 Exa = 1,000 Peta

Currently the fastest supercomputer is “Jaguar” with 2,331 TeraFLOPS. (Top 500 November 2009)

Emulating the Human Brain

The Blue Brain Project has successfully simulated 1 cordical column on the Magerit super computer (100 TeraFLOPS), and human brain has an estimated 1,000,000 columns.

So with 5.2 ExaFlops, one could emulate 52,000 cordial columns. And that’s 1/20th of enough to emulate a human brain.

To emulate the entire human brain you would need approximately 100 ExaFLOPs

That's a lot of FLOPS!

Fore some more notes see: http://www.smartcomputing.com/articles/2002/s1302/39s02/39s02.pdf

Wrap-up

After reading more about the Blue Brain Project, they are emulating a cordial column, as well as analyzing and visualizing that data, which is not necessary for a independent AI.That coupled with the inefficiencies in communication with tele-processing, I suspect that there is a lot of improvement that can be made.

Henry Markham in his recent TED talk, thinks that a human brain could be emulated in 10 years.

I agree with his estimate, using moore's law 10 years seems like an achievable goal.

If you are interested in the subject you may also like Ray Kurzweil's research and books.

Alternatives

More likely a successful AI would be a efficient Emergent-type hive AI. Where independent nodes contribute to a beneficial goal with independent actions. This would not look like a traditional AI to us, the patterns would probably to numerous and vague to track accurately.

This was meant as a thought experiment, if you have any comments please leave them. I will revise this article as I see fit.