![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3223907%2F71515b9453b5607a240de2d7d98fbc33%2FFFT-Time-Frequency-View-540.png?generation=1708463576920394&alt=media\")
This is my first signal processing project so sorry if this sounds dumb. Do spectrograms contain more \"information\" than signals? Do we actually need to convert them or CNN models can learn frequency features from raw signals? What exactly is time to frequency domain conversion?
", "messages": [ { "id": "2660418", "postDate": "02/20/2024 15:36:27", "content": "This is my first signal processing project so sorry if this sounds dumb. Do spectrograms contain more \"information\" than signals? Do we actually need to convert them or CNN models can learn frequency features from raw signals? What exactly is time to frequency domain conversion?
", "rawMarkdown": "This is my first signal processing project so sorry if this sounds dumb. Do spectrograms contain more \"information\" than signals? Do we actually need to convert them or CNN models can learn frequency features from raw signals? What exactly is time to frequency domain conversion?", "votes": null }, { "id": "2660872", "postDate": "02/20/2024 21:21:44", "content": "The raw signal technically has more information within it, and the spectrogram makes the frequency information more accessible to human understanding.
\nNote that the 10-minute spectrograms have information within them that is not present in the raw 50-second EEG signal.
\nI always found an image like this helpful when understanding time to frequency domain conversion:
\n
Simply put, the original signal resides in the time domain, characterized by its amplitude fluctuations over time. The conversion to the frequency domain is achieved by decomposing the signal into constituent sinusoidal waves, each with a distinct frequency, to analyze how their amplitudes—collectively known as the power spectrum—vary across different frequencies.
\nRead more here
", "rawMarkdown": "The raw signal technically has more information within it, and the spectrogram makes the frequency information more accessible to human understanding.\n\nNote that the 10-minute spectrograms have information within them that is not present in the raw 50-second EEG signal.\n\nI always found an image like this helpful when understanding time to frequency domain conversion:\n\n\n\nSimply put, the original signal resides in the time domain, characterized by its amplitude fluctuations over time. The conversion to the frequency domain is achieved by decomposing the signal into constituent sinusoidal waves, each with a distinct frequency, to analyze how their amplitudes—collectively known as the power spectrum—vary across different frequencies.\n\nRead more [here](https://www.nti-audio.com/en/support/know-how/fast-fourier-transform-fft)", "votes": null }, { "id": "2660880", "postDate": "02/20/2024 21:28:08", "content": "I had a lecture about the topic of this competition and this is what I learned you can correct me if I am wrong:
\nRaw EEG signals are voltage measurements recorded from electrodes placed on the scalp, representing the brain's electrical activity over time. They contain detailed temporal information about the brain's electrical activity.
\nAnalyzing raw EEG signals directly can provide insights into rapid changes in brain activity. However, raw EEG signals can be noisy and contain irrelevant information, making it more difficult to extract meaningful patterns directly.
On the other hand, Spectrograms of EEG signals provide a time-frequency representation of the EEG data, showing how the signal's frequency content changes over time. They offer a more compact representation than raw signals, making it easier to visualize and analyze frequency patterns in EEG data.
\nTherefore, by converting EEG signals into spectrograms, we can potentially reduce the impact of noise and focus on more related features.
", "rawMarkdown": "I had a lecture about the topic of this competition and this is what I learned you can correct me if I am wrong:\n\nRaw EEG signals are voltage measurements recorded from electrodes placed on the scalp, representing the brain's electrical activity over time. They contain detailed temporal information about the brain's electrical activity.\nAnalyzing raw EEG signals directly can provide insights into rapid changes in brain activity. However, raw EEG signals can be noisy and contain irrelevant information, making it more difficult to extract meaningful patterns directly.\n\nOn the other hand, Spectrograms of EEG signals provide a time-frequency representation of the EEG data, showing how the signal's frequency content changes over time. They offer a more compact representation than raw signals, making it easier to visualize and analyze frequency patterns in EEG data. \n\nTherefore, by converting EEG signals into spectrograms, we can potentially reduce the impact of noise and focus on more related features.", "votes": null }, { "id": "2661078", "postDate": "02/21/2024 02:33:17", "content": "I think the spectrogram can be regarded as the result of feature extraction of the original signal, and any mathematical transformation will only reduce the information.
", "rawMarkdown": "I think the spectrogram can be regarded as the result of feature extraction of the original signal, and any mathematical transformation will only reduce the information.", "votes": null }, { "id": "2661513", "postDate": "02/21/2024 09:57:55", "content": "This was a good explanation. I think it's similar to windowing in medical imaging or any kind of thresholding that is used for focusing on a certain range of values. Is this operation non-differentiable like thresholding?
", "rawMarkdown": "This was a good explanation. I think it's similar to windowing in medical imaging or any kind of thresholding that is used for focusing on a certain range of values. Is this operation non-differentiable like thresholding?", "votes": null }, { "id": "2661697", "postDate": "02/21/2024 12:57:41", "content": "Thank you! It's interesting to draw parallels between frequency domain conversion and concepts like windowing in medical imaging or thresholding. While these processes share the common goal of isolating and analyzing specific ranges of data for better understanding or processing, their operational foundations differ significantly.
\nTime to frequency domain conversion, especially through the Fourier transform, is a mathematical operation that is fundamentally differentiable with respect to the signal's continuous properties. This transformation is smooth and continuous, allowing for the analysis of frequency components across the entire spectrum.
\nOn the other hand, windowing and thresholding are typically non-differentiable operations. They involve applying a discrete cut-off or selection criterion, which creates a 'jump' or 'step' in the data. This is because thresholding, for example, abruptly changes values based on whether they are above or below a certain threshold, lacking the smooth transition that a differentiable function would have.
", "rawMarkdown": "Thank you! It's interesting to draw parallels between frequency domain conversion and concepts like windowing in medical imaging or thresholding. While these processes share the common goal of isolating and analyzing specific ranges of data for better understanding or processing, their operational foundations differ significantly.\n\nTime to frequency domain conversion, especially through the Fourier transform, is a mathematical operation that is fundamentally differentiable with respect to the signal's continuous properties. This transformation is smooth and continuous, allowing for the analysis of frequency components across the entire spectrum.\n\nOn the other hand, windowing and thresholding are typically non-differentiable operations. They involve applying a discrete cut-off or selection criterion, which creates a 'jump' or 'step' in the data. This is because thresholding, for example, abruptly changes values based on whether they are above or below a certain threshold, lacking the smooth transition that a differentiable function would have.", "votes": null } ], "comments": [ { "id": 2660872, "author_name": "seanbearden", "author_url": "", "post_date": "02/20/2024 21:21:44", "content": "The raw signal technically has more information within it, and the spectrogram makes the frequency information more accessible to human understanding.
\nNote that the 10-minute spectrograms have information within them that is not present in the raw 50-second EEG signal.
\nI always found an image like this helpful when understanding time to frequency domain conversion:
\n
Simply put, the original signal resides in the time domain, characterized by its amplitude fluctuations over time. The conversion to the frequency domain is achieved by decomposing the signal into constituent sinusoidal waves, each with a distinct frequency, to analyze how their amplitudes—collectively known as the power spectrum—vary across different frequencies.
\nRead more here
", "votes": null, "replies": [ { "id": 2661513, "author_name": "gunesevitan", "author_url": "", "post_date": "02/21/2024 09:57:55", "content": "This was a good explanation. I think it's similar to windowing in medical imaging or any kind of thresholding that is used for focusing on a certain range of values. Is this operation non-differentiable like thresholding?
", "votes": null, "replies": [ { "id": 2661697, "author_name": "seanbearden", "author_url": "", "post_date": "02/21/2024 12:57:41", "content": "Thank you! It's interesting to draw parallels between frequency domain conversion and concepts like windowing in medical imaging or thresholding. While these processes share the common goal of isolating and analyzing specific ranges of data for better understanding or processing, their operational foundations differ significantly.
\nTime to frequency domain conversion, especially through the Fourier transform, is a mathematical operation that is fundamentally differentiable with respect to the signal's continuous properties. This transformation is smooth and continuous, allowing for the analysis of frequency components across the entire spectrum.
\nOn the other hand, windowing and thresholding are typically non-differentiable operations. They involve applying a discrete cut-off or selection criterion, which creates a 'jump' or 'step' in the data. This is because thresholding, for example, abruptly changes values based on whether they are above or below a certain threshold, lacking the smooth transition that a differentiable function would have.
", "votes": null, "replies": [] } ] } ] }, { "id": 2660880, "author_name": "feidawei", "author_url": "", "post_date": "02/20/2024 21:28:08", "content": "I had a lecture about the topic of this competition and this is what I learned you can correct me if I am wrong:
\nRaw EEG signals are voltage measurements recorded from electrodes placed on the scalp, representing the brain's electrical activity over time. They contain detailed temporal information about the brain's electrical activity.
\nAnalyzing raw EEG signals directly can provide insights into rapid changes in brain activity. However, raw EEG signals can be noisy and contain irrelevant information, making it more difficult to extract meaningful patterns directly.
On the other hand, Spectrograms of EEG signals provide a time-frequency representation of the EEG data, showing how the signal's frequency content changes over time. They offer a more compact representation than raw signals, making it easier to visualize and analyze frequency patterns in EEG data.
\nTherefore, by converting EEG signals into spectrograms, we can potentially reduce the impact of noise and focus on more related features.
", "votes": null, "replies": [] }, { "id": 2661078, "author_name": "sunyuri", "author_url": "", "post_date": "02/21/2024 02:33:17", "content": "I think the spectrogram can be regarded as the result of feature extraction of the original signal, and any mathematical transformation will only reduce the information.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2660418": "This is my first signal processing project so sorry if this sounds dumb. Do spectrograms contain more \"information\" than signals? Do we actually need to convert them or CNN models can learn frequency features from raw signals? What exactly is time to frequency domain conversion?", "2660872": "The raw signal technically has more information within it, and the spectrogram makes the frequency information more accessible to human understanding.\n\nNote that the 10-minute spectrograms have information within them that is not present in the raw 50-second EEG signal.\n\nI always found an image like this helpful when understanding time to frequency domain conversion:\n\n\n\nSimply put, the original signal resides in the time domain, characterized by its amplitude fluctuations over time. The conversion to the frequency domain is achieved by decomposing the signal into constituent sinusoidal waves, each with a distinct frequency, to analyze how their amplitudes—collectively known as the power spectrum—vary across different frequencies.\n\nRead more [here](https://www.nti-audio.com/en/support/know-how/fast-fourier-transform-fft)", "2660880": "I had a lecture about the topic of this competition and this is what I learned you can correct me if I am wrong:\n\nRaw EEG signals are voltage measurements recorded from electrodes placed on the scalp, representing the brain's electrical activity over time. They contain detailed temporal information about the brain's electrical activity.\nAnalyzing raw EEG signals directly can provide insights into rapid changes in brain activity. However, raw EEG signals can be noisy and contain irrelevant information, making it more difficult to extract meaningful patterns directly.\n\nOn the other hand, Spectrograms of EEG signals provide a time-frequency representation of the EEG data, showing how the signal's frequency content changes over time. They offer a more compact representation than raw signals, making it easier to visualize and analyze frequency patterns in EEG data. \n\nTherefore, by converting EEG signals into spectrograms, we can potentially reduce the impact of noise and focus on more related features.", "2661078": "I think the spectrogram can be regarded as the result of feature extraction of the original signal, and any mathematical transformation will only reduce the information.", "2661513": "This was a good explanation. I think it's similar to windowing in medical imaging or any kind of thresholding that is used for focusing on a certain range of values. Is this operation non-differentiable like thresholding?", "2661697": "Thank you! It's interesting to draw parallels between frequency domain conversion and concepts like windowing in medical imaging or thresholding. While these processes share the common goal of isolating and analyzing specific ranges of data for better understanding or processing, their operational foundations differ significantly.\n\nTime to frequency domain conversion, especially through the Fourier transform, is a mathematical operation that is fundamentally differentiable with respect to the signal's continuous properties. This transformation is smooth and continuous, allowing for the analysis of frequency components across the entire spectrum.\n\nOn the other hand, windowing and thresholding are typically non-differentiable operations. They involve applying a discrete cut-off or selection criterion, which creates a 'jump' or 'step' in the data. This is because thresholding, for example, abruptly changes values based on whether they are above or below a certain threshold, lacking the smooth transition that a differentiable function would have." }, "source": "meta" }