5 ms·

Could we have something like an AM broadcast, but then the receiver “upsampling” the data? Maybe using some ML? For example, if you have an AM channel that alw

by nico 2y ago

Could we have something like an AM broadcast, but then the receiver “upsampling” the data? Maybe using some ML? For example, if you have an AM channel that always transmits some song from a list, then you might not even need to receive the whole song. Or maybe you can have a model that’s good at upsampling those types of songs/data?

mlyle 2y ago
This is pretty much what audio compression, like a MP3 file, does: it figures out how to encode perceptually-identical (or close) stuff more densely and remove redundant information. Every time you listen to "HD Radio" you're doing exactly this.
- vlovich123 2y ago
  Yes, but modern ML techniques (particularly with neural networks) are much more impressive at recovering signal with even less information than things like MP3. An example of this in the visual domain is upscaling. So it’s not a bad question at all.
  - mlyle 2y ago
    Sure-- as we've discussed endlessly here: compression and ML are really very close cousins. I never said it was a bad question.
    - keithalewis 2y ago
      You can even win a prize for that https://en.m.wikipedia.org/wiki/Hutter_Prize https://en.m.wikipedia.org/wiki/Hutter_Prize.
      - vlovich123 2y ago
        I extremely dislike the Hutter Prize. I think it's generally agreed that intelligence compresses the model better & therefore it's reasonable to assume that the better the compression, the more intelligence you have (at least for now where we are with AI, this seems to be a roughly reasonable assumption). My problem with the Hutter prize is that it makes the leap to say that lossless compression is this when everything about AI is choosing how much error you're willing to tolerate. In other words, intelligence is linked to lossy compression, not lossless - a competition for lossless compression doesn't do anything more than get us better lossless compressors / decompressors.
YZF 2y ago
AM broadcast uses a "band" around the carrier. Whatever tricks you use Shannon's law is going to determine the upper limit of how much data you can carry which depends on the bandwidth and the signal to noise ratio. ML makes no difference from a theoretical perspective. Compression just reduces the bandwidth you'll need assuming your data compresses (with whatever method you choose, ML can be part of that)...
- nico 2y ago
  I guess my usual experience is that AM transmissions sound pretty horrible, especially compared to FM So, my question is more like, can we just upgrade the equipment and keep the channels? And can AM be used more like for signal/indexing rather than full data dump? (eg. instead of transmitting the whole song, transmitting just an id of the song)
  - vlovich123 2y ago
    AM sounds bad because it’s amplitude modulated. So variation in amplitude (i.e. signal strength) results in noise when decoding. FM however is not sensitive to this issue because its frequency modulated - slight variations in signal strength have no effect on the encoded signal. And yes, we can always upgrade the equipment. That’s what we did when we transitioned TV to digital signals and FM to digital FM - we reserved some bandwidth for over-the-air while freeing up the rest to be used for other purposes. But it’s an expensive proposition and one that happens rarely (in fact those are the only two times I’m aware of it, they happened at basically the same time, took forever, & there was insane opposition to it).
  - YZF 2y ago
    The limit is Shannon's law which depends on bandwidth and signal to noise. A more advanced modulation technique can get more "information" through up to that limit. A song for this matter can just be considered information as well. So the answer is almost certainly yes, one can get more out of existing channels by upgrading the equipment assuming the current equipment doesn't approach the theoretical limit which is almost certainly the case for this "old" technology we're talking about. An FM radio channel has a bandwidth of 200Khz. AM radio has a bandwidth of of 10Khz. That's a big part of the reason why AM sounds horrible compared to FM. The way noise impacts AM and FM is also different, especially given human perception. IIRC both these methods were picked for practical reasons not necessarily any optimal criteria. AM is fairly crude, there's only so much you can change the amplitude for any practical purpose. Changing the frequency makes it easier to use more bandwidth so it was a natural evolution. Depending on how you're modulating the amplitude (the naive old school radio IIRC is just using the signal you're transmitting to modulate it) you can make the bandwidth larger or smaller. Random e.g. something like Ethernet on copper can be viewed as a very very fast, high bandwidth, AM. All these things boil down to the same Shannon law just with somewhat different efficiencies depending on how well you can control the frequency domain. This problem of cramming the most bits into a given channel is something people have been working on for a long time. Some coverage of this here: https://en.wikipedia.org/wiki/Modulation https://en.wikipedia.org/wiki/Modulation That's how we got from 300bps modems to 56k modems over the same phone lines way back and how we push more out of radio spectrum today.

by nico 2y ago

mlyle 2y ago

This is pretty much what audio compression, like a MP3 file, does: it figures out how to encode perceptually-identical (or close) stuff more densely and remove redundant information. Every time you listen to "HD Radio" you're doing exactly this.

vlovich123 2y ago
Yes, but modern ML techniques (particularly with neural networks) are much more impressive at recovering signal with even less information than things like MP3. An example of this in the visual domain is upscaling. So it’s not a bad question at all.
- mlyle 2y ago
  Sure-- as we've discussed endlessly here: compression and ML are really very close cousins. I never said it was a bad question.
  - keithalewis 2y ago
    You can even win a prize for that https://en.m.wikipedia.org/wiki/Hutter_Prize https://en.m.wikipedia.org/wiki/Hutter_Prize.
    - vlovich123 2y ago
      I extremely dislike the Hutter Prize. I think it's generally agreed that intelligence compresses the model better & therefore it's reasonable to assume that the better the compression, the more intelligence you have (at least for now where we are with AI, this seems to be a roughly reasonable assumption). My problem with the Hutter prize is that it makes the leap to say that lossless compression is this when everything about AI is choosing how much error you're willing to tolerate. In other words, intelligence is linked to lossy compression, not lossless - a competition for lossless compression doesn't do anything more than get us better lossless compressors / decompressors.

YZF 2y ago

AM broadcast uses a "band" around the carrier. Whatever tricks you use Shannon's law is going to determine the upper limit of how much data you can carry which depends on the bandwidth and the signal to noise ratio. ML makes no difference from a theoretical perspective. Compression just reduces the bandwidth you'll need assuming your data compresses (with whatever method you choose, ML can be part of that)...

nico 2y ago
I guess my usual experience is that AM transmissions sound pretty horrible, especially compared to FM So, my question is more like, can we just upgrade the equipment and keep the channels? And can AM be used more like for signal/indexing rather than full data dump? (eg. instead of transmitting the whole song, transmitting just an id of the song)
- vlovich123 2y ago
  AM sounds bad because it’s amplitude modulated. So variation in amplitude (i.e. signal strength) results in noise when decoding. FM however is not sensitive to this issue because its frequency modulated - slight variations in signal strength have no effect on the encoded signal. And yes, we can always upgrade the equipment. That’s what we did when we transitioned TV to digital signals and FM to digital FM - we reserved some bandwidth for over-the-air while freeing up the rest to be used for other purposes. But it’s an expensive proposition and one that happens rarely (in fact those are the only two times I’m aware of it, they happened at basically the same time, took forever, & there was insane opposition to it).
- YZF 2y ago
  The limit is Shannon's law which depends on bandwidth and signal to noise. A more advanced modulation technique can get more "information" through up to that limit. A song for this matter can just be considered information as well. So the answer is almost certainly yes, one can get more out of existing channels by upgrading the equipment assuming the current equipment doesn't approach the theoretical limit which is almost certainly the case for this "old" technology we're talking about. An FM radio channel has a bandwidth of 200Khz. AM radio has a bandwidth of of 10Khz. That's a big part of the reason why AM sounds horrible compared to FM. The way noise impacts AM and FM is also different, especially given human perception. IIRC both these methods were picked for practical reasons not necessarily any optimal criteria. AM is fairly crude, there's only so much you can change the amplitude for any practical purpose. Changing the frequency makes it easier to use more bandwidth so it was a natural evolution. Depending on how you're modulating the amplitude (the naive old school radio IIRC is just using the signal you're transmitting to modulate it) you can make the bandwidth larger or smaller. Random e.g. something like Ethernet on copper can be viewed as a very very fast, high bandwidth, AM. All these things boil down to the same Shannon law just with somewhat different efficiencies depending on how well you can control the frequency domain. This problem of cramming the most bits into a given channel is something people have been working on for a long time. Some coverage of this here: https://en.wikipedia.org/wiki/Modulation https://en.wikipedia.org/wiki/Modulation That's how we got from 300bps modems to 56k modems over the same phone lines way back and how we push more out of radio spectrum today.