ad: ProAudio-1

Can neural networks beat humans at decoding Morse code?

Discussion in 'General Technical Questions and Answers' started by M0AGP, Jan 20, 2021.

ad: L-HROutlet
ad: l-rl
ad: Left-3
ad: L-MFJ
ad: Subscribe
ad: QSOToday-1
ad: abrind-2
ad: Left-2
  1. M0AGP

    M0AGP Ham Member QRZ Page

    Neural networks can do amazing things like drive cars safely and spot cats in videos. There are some academic papers I can’t understand where people seem to try this - surely neural networks should be able to be trained to copy CW better than humans, no?

    Personally I like head copy, but I am just interested intellectually.

    (Moderator please move this to a better location if there is one - sorry, I couldn’t see an obvious place for it. The CW forum might hate the idea and the software forums didn’t seem like the right places either.)
     
  2. AG6QR

    AG6QR Premium Subscriber QRZ Page

    I don't know enough about neural networks to say what they can and can't do.

    I do know about conventional procedural programming, and I believe that, once you teach the computer to reliably know when the key is closed or not, the rest of the decoding isn't so hard. Human ears can naturally recognize tones, and to us, a static crash sounds nothing like a CW tone. But I believe many CW decoders seem to do some crude filtering and then just detect volume of sound, and are therefore easily distracted by noise and static.

    The software that runs the Reverse Beacon Network seems to have very good ability to decode callsigns across wide swaths of the bands at once. I'd say it's already better at its task than any human I know. Though I know humans who can decode one transmission at a time very well, I don't know anyone who can copy a whole band at once. I suspect they're using some type of Fast Fourier Transform to detect the tones.

    Most CW decoders are junk. But I find it hard to believe that they have to be so bad.
     
    M0AGP likes this.
  3. PU2OZT

    PU2OZT Ham Member QRZ Page

    AG5DB, N0TZU and M0AGP like this.
  4. GNUUSER

    GNUUSER QRZ Member


    electronic neural networks are light years behind the complexity of the human neural network.
    for all their speed they only have a limited acumen to access.
    cw decoders have a limited amount of data to rely on. they cannot differentiate speed differences adequately enough to accurately decode the signal,
    you set the speed for say 10 words a minute (because that's what you can send) and someone replying keys at 12 1/2 wpm and it may be outside the parameter of the decoder subroutine, ( you get some message but also a lot of misc. characters and letters)

    consider this analogy you as the computer have to look for a yellow marble in a blue box with 5 other different colored marbles and say the box is about the size of a 3 pound coffee can.
    you can find that yellow marble pretty darn fast because you don't have to look very far.
    the same is said about autopilot systems that can navigate your car.
    they have limited sensors, small cache of subroutines, but no actual conscious decision. only the best and highest probability result of an algorithm.
    computers are fast But they by far less intelligent then a urinal.
    any computer no matter how complex can only run programs put into it.

    now the human mind however can be likened to a blue box the size of lake superior!
    now the yellow marble sticks out like a sore thumb, But it would still be difficult to see, With some detailed searching you would eventually find it.
    now listening to morse code a trained operator can pick up and decode information at a phenomenal speed,
    unless your completely stone deaf the human ear can detect an incredible range of information and subtle changes in pitch and or tone.
    no computer has the ability to do that.
    so in this instance just how fast is your mind searching the lake superior box for the yellow marble and finding it?
     
    M0AGP likes this.
  5. N0TZU

    N0TZU Platinum Subscriber Platinum Subscriber QRZ Page

    Great idea, and it should be achievable right now. Such audio artificial intelligence is in our lives already - speech recognition systems used in almost all phone call centers and in dictation systems. Adapting this for Morse code should be doable.

    At least for vision systems, the hardware and software is already on the market at reasonable prices, and much is open source. My son’s company makes small, relatively inexpensive, stand alone 3D video vision systems that incorporate amazingly sophisticated neural network AI detection of all sorts of things, and they can be readily trained for specialized applications. I played with one of their early models and it was quite astonishing how well it worked to identify and track objects and people in real time.
     
    M0AGP likes this.
  6. M0AGP

    M0AGP Ham Member QRZ Page

    There will be of course certain things computers can do better in copying CW: for example they will be able to copy 1000 words per minute. I guess I was thinking about weak signal work, where it is said that very skilled humans can copy at “-15dB” which is supposed to be on the same S/N scale as digital modes.

    You can imagine creating increasingly challenging training datasets for a network, starting with perfect CW at one speed, once trained, add varying speed. Once trained add noise, then even more noise, etc.

    I must admit I haven’t read the interesting looking articles on OZT’s reply - thanks for posting that!

    I can see what I’ll be doing on lunch break!
     
    PU2OZT likes this.
  7. PU2OZT

    PU2OZT Ham Member QRZ Page

  8. K2CAJ

    K2CAJ XML Subscriber QRZ Page

    Good question. Because the code itself is so simple and well-defined (and designed for unambiguous communication,) it sounds like the real problem is detecting whether a weak signal exists, and tracking the pitch and tone.

    As an EE, my gut feeling is that you might want to try a "classical" method first, because picking out a single tone in white noise is such a classical problem almost tailor-made for classical methods. At they very least, anyone attempting deep learning would have to pit it against classical parameter estimation to get any sense of improved performance.
     
    KA0HCP likes this.
  9. KA5IPF

    KA5IPF Ham Member QRZ Page

    If you think AI is getting there turn on closed captioning on your TV and see how it butchers the english language.
     
    N0NC, KP4SX and K2CAJ like this.
  10. N0TZU

    N0TZU Platinum Subscriber Platinum Subscriber QRZ Page

    I think the major issue with algorithms is variable timing on every element and lack of a clock and synchronization characters.
     
    K6CLS likes this.

Share This Page

ad: M2Ant-1