Hacker News Reader

Comment by wcoenen | Hacker News Reader

The top voted answer is assuming a neuron firing/not firing (once per refractory period) can be counted as a bit for the purpose of calculating bandwidth.

But I think the peripheral nerve system uses firing frequency to encode intensity, so I'm not sure you can really equate 1 firing with one bit.

For example, over 15 refractory periods there could be anywhere between 0 and 15 firings, thereby encoding 16 different possible intensities. That would effectively be only a 4 bit message per 15 periods, not 15 bit per 15 periods.

View on HN

The classic goodput vs throughput debate. E.g. Gigabit Ethernet is most commonly actually physically transmitted at 1.25 gigabit/s because of 8b10b coding. Ethernet requires interpacket gaps, and the packetization itself changes overhead. Then of course you have other protocol overheads underneath you could consider, and how much overhead the data you were transmitting has due to encoding of it itself.

In the end there is really "bandwidth" which is how much symbol "space" is available and various levels of "goodput" which is the rate of whatever you're calling useful data.

baud vs. bps

Other things like unusable symbols (e.g. IFG added to above) make it more nuanced than that.

> Gigabit Ethernet is most commonly actually physically transmitted at 1.25 gigabit/s because of 8b10b coding.

Fiber often uses 8b/10b. Over copper it's much more complicated.

USB 3.0 is an example of simple 8b/10b. 4Gbps data, sent at 5Gbps.

Ethernet over twisted pair, well...

10mbit Ethernet turns each bit into either a 01 or a 10 and sends those on the line.

100mbit Ethernet applies 4b5 encoding to get a 125MHz signal, then it uses this to drive an encoding called MLT-3. In MLT-3, the final output cycles through -1, 0, +1, 0 over and over, and each bit decides whether it advances or doesn't advance.

Gigabit Ethernet does a complicated encoding to turn batches of 8 bits into 4 separate voltages from -2 to +2 (so 5 levels). It then sends these simultaneously over all four wire pairs at a grand total of 125MHz.

Timing is important with frequency encoding intensity.

So, max frequency is encoding 100% as fast as possible, but this means it’s possible to have effectively 15x100% or 14 at 99.99%, 99.97, 99.98 ... or really any combination depending on when each pulse specifically ends.

This means you need to decode how accurate the timing information is to figure out the bandwidth but each pulse encodes a floating point number not a bit.

PS: For the most ridiculous upper limit. If a neuron could fire 500 times+ a second with a 1ms refractory period each encoding a number between 0 to ~10^40 or ~2^130 as the plank unit of time beyond the minimum. That’s ~50,000+ bits per neuron per second.

Good point and nice estimation! Measured bits/spike (e.g. in cricket sensory neurons) tops out at 3.2 bits/spike [1]. Most sensory neurons across all species top out at ~300 bits/second.

So if we assume a similar bit rate for spine neurons, we get 7.5 GBps, not 16.625 GBps.

[1] : https://www.nature.com/articles/nn1199_947

Though, it probably never reaches even remotely full capacity except in rare events like seizures or when jumping into cold water. Also, there is a fair amount of redundancy in neural information processing due to the noisy computation substrate. I would not be surprised if the channel capacity is something like 100 Mb/s, but transmission rates are less than 10 Mb/s most of the time.

You're fine tuning a parameter of a rough-order-of-magnitude estimation - I think each of the parameters can be discussed in the same fashion, your best hope is that errors cancel out and you're finding a result that is only a few factors out

So, sort of like 4-QAM then?

Reply on news.ycombinator.com