‘Insufficient facts always invite danger.’
Spock in Star Trek TOS Episode “Space Seed”
As most CSP Blog readers likely know, I’ve performed detailed critical analyses (one, two, three, and four) of the modulation-recognition datasets put forth publicly by DeepSig in 2016-2018. These datasets are associated with some of their published or arxiv.org papers, such as The Literature [R138], which I also reviewed here.
My conclusion is that the DeepSig datasets are as flawed as the DeepSig papers–it was the highly flawed nature of the papers that got me started down the critical-review path in the first place.
A reader recently alerted me to a change in the Datasets page at deepsig.ai that may indicate they are listening to critics. Let’s take a look and see if there is anything more to say.
Danger Will Robinson! Non-technical post approaching!
When I was a wee engineer, I’d sometimes clash with other engineers that sneered at technical approaches that didn’t set up a linear-algebraic optimization problem as the first step. Never mind that I’ve been relentlessly focused on single-sensor problems, rather than array-processing problems, and so the naturalness of the linear-algebraic mathematical setting was debatable–however there were still ways to fashion matrices and compute those lovely eigenvalues. The real issue wasn’t the dimensionality of the data model, it was that I didn’t have a handy crank I could turn and pop out a provably optimal solution to the posed problem. Therefore I could be safely ignored. And if nobody could actually write down an optimization problem for, say, general radio-frequency scene analysis, then that problem just wasn’t worth pursuing.
Those critical engineers worship at the altar of optimality. Time for another rant.
By the pricking of my thumbs, something wicked this way comes …
I attended a conference on dynamic spectrum access in 2017 and participated in a session on automatic modulation recognition. The session was connected to a live competition within the conference where participants would attempt to apply their modulation-recognition system to signals transmitted in the conference center by the conference organizers. Like a grand modulation-recognition challenge but confined to the temporal, spectral, and spatial constraints imposed by the short-duration conference.
What I didn’t know going in was the level of frustration on the part of the machine-learner organizers regarding the seemingly inability of signal-processing and machine-learning researchers to solve the radio-frequency scene analysis problem once and for all. The basic attitude was ‘if the image-processors can have the AlexNet image-recognition solution, and thereby abandon their decades-long attempt at developing serious mathematics-based image-processing theory and practice, why haven’t we solved the RFSA problem yet?’
The next step in dataset complexity at the CSP Blog: cochannel signals.
I’ve developed another dataset for use in assessing modulation-recognition algorithms (machine-learning-based or otherwise) that is more complex than the original sets I posted for the ML Challenge (CSPB.ML.2018 and CSPB.ML.2022). Half of the new dataset consists of one signal in noise and the other half consists of two signals in noise. In most cases the two signals overlap spectrally, which is a signal condition called cochannel interference.
How can we train a neural network to make use of both IQ data samples and CSP features in the context of weak-signal detection?
I’ve been working with some colleagues at Northeastern University (NEU) in Boston, MA, on ways to combine CSP with machine learning. The work I’m doing with Old Dominion University is focused on basic modulation recognition using neural networks and, in particular, the generalization (dataset-shift) problem that is pervasive in deep learning with convolution neural networks. In contrast, the NEU work is focused on specific signal detection and classification problems and looks at how to use multiple disparate data types as inputs to neural-networks; inputs such as complex-valued samples (IQ data) as well as carefully selected components of spectral correlation and spectral coherence surfaces.
My NEU colleagues and I will be publishing a rather lengthy conference paper on a new multi-input-data neural-network approach called ICARUS at InfoCom 2023 this May (My Papers ). You can get a copy of the pre-publication version here or on arxiv.org.
Neural networks with CSP-feature inputs DO generalize in the modulation-recognition problem setting.
In some recently published papers (My Papers [50,51]), my ODU colleagues and I showed that convolutional neural networks and capsule networks do not generalize well when their inputs are complex-valued data samples, commonly referred to as simply IQ samples, or as raw IQ samples by machine learners.(Unclear why the adjective ‘raw’ is often used as it adds nothing to the meaning. If I just say Hey, pass me those IQ samples, would ya?, do you think maybe he means the processed ones? How about raw-I-mean–seriously-man–I-did-not-touch-those-numbers-OK? IQ samples? All-natural vegan unprocessed no-GMO organic IQ samples?Uncooked IQ samples?) Moreover, the capsule networks typically outperform the convolutional networks.
In a new paper (MILCOM 2022: My Papers ; arxiv.org version), my colleagues and I continue this line of research by including cyclic cumulants as the inputs to convolutional and capsule networks. We find that capsule networks outperform convolutional networks and that convolutional networks trained on cyclic cumulants outperform convolutional networks trained on IQ samples. We also find that both convolutional and capsule networks trained on cyclic cumulants generalize perfectly well between datasets that have different (disjoint) probability density functions governing their carrier frequency offset parameters.
That is, convolutional networks do better recognition with cyclic cumulants and generalize very well with cyclic cumulants.
So why don’t neural networks ever ‘learn’ cyclic cumulants with IQ data at the input?
The majority of the software and analysis work is performed by the first author, John Snoap, with an assist on capsule networks by James Latshaw. I created the datasets we used (available here on the CSP Blog [see below]) and helped with the blind parameter estimation. Professor Popescu guided us all and contributed substantially to the writing.
Another brick in the wall, another drop in the bucket, another windmill on the horizon …
Let’s talk more about The Cult. No, I don’t mean She Sells Sanctuary, for which I do have considerable nostalgic fondness. I mean the Cult(ure) of Machine Learning in RF communications and signal processing. Or perhaps it is more of an epistemic bubble where there are The Things That Must Be Said and The Unmentionables in every paper and a style of research that is strictly adhered to but that, sadly, produces mostly error and promotes mostly hype. So we have shibboleths, taboos, and norms to deal with inside the bubble.
Time to get on my high horse. She’s a good horse named Ravager and she needs some exercise. So I’m going to strap on my claymore, mount Ravager, and go for a ride. Or am I merely tilting at windmills?
Let’s take a close look at another paper on machine learning for modulation recognition. It uses, uncritically, the DeepSig RML 2016 datasets. And the world and the world, the world drags me down…
Back in 2018 I posted a dataset consisting of 112,000 I/Q data files, 32,768 samples in length each, as a part of a challenge to machine learners who had been making strong claims of superiority over signal processing in the area of automatic modulation recognition. One part of the challenge was modulation recognition involving eight digital modulation types, and the other was estimating the carrier frequency offset. That dataset is described here, and I’d like to refer to it as CSPB.ML.2018.
Then in 2022 I posted a companion dataset to CSPB.ML.2018 called CSPB.ML.2022. This new dataset uses the same eight modulation types, similar ranges of SNR, pulse type, and symbol rate, but the random variable that governs the carrier frequency offset is different with respect to the random variable in CSPB.ML.2018. The purpose of the CSPB.ML.2022 dataset is to facilitate studies of the dataset-shift, or generalization, problem in machine learning.
Throughout the past couple of years I’ve been working with some graduate students and a professor at Old Dominion University on merging machine learning and signal processing for problems involving RF signal analysis, such as modulation recognition. We are starting to publish a sequence of papers that describe our efforts. I briefly describe the results of one such paper, My Papers , in this post.
While reading a book on string theory for lay readers, I did a double take…
I don’t know why I haven’t read any of Lee Smolin’s physics books prior to this year, but I haven’t. Maybe blame my obsession with Sean Carroll. In any case, I’ve been reading The Trouble with Physics (The Literature [R175]), which is about string theory and string theorists. Smolin finds it troubling that the string theorist subculture in physics shows some signs of groupthink and authoritarianism. Perhaps elder worship too.
I came across this list of attributes, conceived by Smolin, of the ‘sociology’ of the string-theorist contingent:
The softwarization of engineering continues apace…
I keep seeing people write things like “a major disadvantage of the technique for X is that it requires substantial domain expertise.” Let’s look at a recent good paper that makes many such remarks and try to understand what it could mean, and if having or getting domain expertise is actually a bad thing. Spoiler: It isn’t.
The paper under the spotlight is The Literature [R174], “Interference Suppression Using Deep Learning: Current Approaches and Open Challenges,” published for the nonce on arxiv.org. I’m not calling this post a “Comments On …” post, because once I extract the (many) quotes about domain expertise, I’m leaving the paper alone. The paper is a good paper and I expect it to be especially useful for current graduate students looking to make a contribution in the technical area where machine learning and RF signal processing overlap. I especially like Figure 1 and the various Tables.
Another RF-signal dataset to help push along our R&D on modulation recognition.
Update February 2023: A third dataset has been posted to the CSP Blog: CSPB.ML.2023. It features cochannel signals.
Update January 2023: I’m going to put Challenger results in the Comments. I’ve received a Challenger’s decisions and scored them in January 2023. See below.
In this post I provide a second dataset for the Machine-Learning Challenge I issued in 2018 (CSPB.ML.2018). This dataset is similar to the original dataset, but possesses a key difference in that the probability distribution of the carrier-frequency offset parameter, viewed as a random variable, is not the same, but is still realistic.
Blog Note: By WordPress’ count, this is the 100th post on the CSP Blog. Together with a handful of pages (like My Papers and The Literature), these hundred posts have resulted in about 250,000 page views. That’s an average of 2,500 page views per post. However, the variance of the per-post pageviews is quite large. The most popular is The Spectral Correlation Function (> 16,000) while the post More on Pure and Impure Sinewaves, from the same era, has only 316 views. A big Thanks to all my readers!!
We take a quick look at a fourth DeepSig dataset called 2016.04C.multisnr.tar.bz2 in the context of the data-shift problem in machine learning.
And if we get this right,
We’re gonna teach ’em how to say
You and I.
Lin-Manuel Miranda, “One Last Time,” Hamilton
I didn’t expect to have to do this, but I am going to analyze yet another DeepSig dataset. One last time. This one is called 2016.04C.multisnr.tar.bz2, and is described thusly on the DeepSig website:
I’ve analyzed the 2018 dataset here, the RML2016.10b.tar.bz2 dataset here, and the RML2016.10a.tar.bz2 dataset here.
Now I’ve come across a manuscript-in-review in which both the RML2016.10a and RML2016.04c data sets are used. The idea is that these two datasets represent two sufficiently distinct datasets so that they are good candidates for use in a data-shift study involving trained neural-network modulation-recognition systems.
The data-shift problem is, as one researcher puts it:
Data shift or data drift, concept shift, changing environments, data fractures are all similar terms that describe the same phenomenon: the different distribution of data between train and test sets
Does the use of ‘total SNR’ mislead when the fractional bandwidth is very small? What constitutes ‘weak-signal processing?’
Or maybe “Comments on” here should be “Questions on.”
In a recent paper in EURASIP Journal on Advances in Signal Processing (The Literature [R165]), the authors tackle the problem of machine-learning-based modulation recognition for highly oversampled rectangular-pulse digital signals. They don’t use the DeepSig datasets (one, two, three, four), but their dataset description and use of ‘signal-to-noise ratio’ leaves a lot to be desired. Let’s take a brief look. See if you agree with me that the touting of their results as evidence that they can reliably classify signals with ‘SNRs of dB’ is unwarranted and misleading.
Another post-publication review of a paper that is weak on the ‘RF’ in RF machine learning.
Let’s take a look at a recently published paper (The Literature [R148]) on machine-learning-based modulation-recognition to get a data point on how some electrical engineers–these are more on the side of computer science I believe–use mathematics when they turn to radio-frequency problems. You can guess it isn’t pretty, and that I’m not here to exalt their acumen.
The Machine Learners think that their “feature engineering” (rooting around in voluminous data) is the same as “features” in mathematically derived signal-processing algorithms. I take a lighthearted look.
One of the things the machine learners never tire of saying is that their neural-network approach to classification is superior to previous methods because, in part, those older methods use hand-crafted features. They put it in different ways, but somewhere in the introductory section of a machine-learning modulation-recognition paper (ML/MR), you’ll likely see the claim. You can look through the ML/MR papers I’ve cited in The Literature ([R133]-[R146]) if you are curious, but I’ll extract a couple here just to illustrate the idea.
DeepSig’s datasets are popular in the machine-learning modulation-recognition community, and in that community there are many claims that the deep neural networks are vastly outperforming any expertly hand-crafted tired old conventional method you care to name (none are usually named though). So I’ve been looking under the hood at these datasets to see what the machine learners think of as high-quality inputs that lead to disruptive upending of the sclerotic mod-rec establishment. In previous posts, I’ve looked at two of the most popular DeepSig datasets from 2016 (here and here). In this post, we’ll look at one more and I will then try to get back to the CSP posts.
Let’s take a look at one more DeepSig dataset: 2018.01.OSC.0001_1024x2M.h5.tar.gz.
The second DeepSig data set I analyze: SNR problems and strange PSDs.
I presented an analysis of one of DeepSig’s earlier modulation-recognition datasets (RML2016.10a.tar.bz2) in the post on All BPSK Signals. There we saw several flaws in the dataset as well as curiosities. Most notably, the signals in the dataset labeled as analog amplitude-modulated single sideband (AM-SSB) were absent: these signals were only noise. DeepSig has several other datasets on offer at the time of this writing:
In this post, I’ll present a few thoughts and results for the “Larger Version” of RML2016.10a.tar.bz2, which is called RML2016.10b.tar.bz2. This is a good post to offer because it is coherent with the first RML post, but also because more papers are being published that use the RML 10b dataset, and of course more such papers are in review. Maybe the offered analysis here will help reviewers to better understand and critique the machine-learning papers. The latter do not ever contain any side analysis or validation of the RML datasets (let me know if you find one that does in the Comments below), so we can’t rely on the machine learners to assess their inputs. (Update: I analyze a third DeepSig dataset here. And a fourth and final one here.)