1
$\begingroup$

This is not a technical question but can have a yes/no answer, and a why statement. My understanding about digital signal processing (DSP) is that it allows to take real-world analog signals (1D, 2D, etc...), sample them to create digital representations on the computer, and either way: extract useful information from filters or modify the signal to take it back to the analog world (using a digital-to-analog converter if you like). There is an overlap with other disciplines like pattern classification or statistics as DSP implements dimensionality reduction techniques that make easier to classify samples; or with embedded systems, that have DSP ICs. I am not considering analog signal processing which is another topic for a question.

There a talk explaining that there is an "identity crisis" in DSP nowadays with all the machines learning (ML) methods because they can process raw data without having to apply the carefully crafted DSP knowledge (let's say, using a CNN instead of extracting feature vectors). Of course, sacrificing computational power, time tunning hyperparameters and or the effort to create enough training data. But people don't care as long as they get the processing they want and don't have CPU, power, or memory constraints. This also reminds me this article pointing that DSP roles are not longer the same they used to be 25 years ago.

I have observed certain similarity between computer vision (CV) and data science roles with DSP, but I am still trying to convince my self that someone can really make some DSP working in them. On a CV role you are in charge of dealing with 2D+ signals and perform tasks such as face recognition or pose estimation (which has historically been a DSP task) but the problem is that the majority of the work is to reuse components-out-of-the-shelf (COTS), like openCV, to specific areas, without having to demonstrate full understanding of the yielding techniques (of course, I am talking about someone that hasn't to work in a lab where some research work needs to be done). And about data science, I know that they use a lot of ML, data visualization, statistics, business logic, A/B testing; and have to use the data gathered by data engineers, but they do not use DSP-core techniques like sampling and interpolation or FIR/IIR/adaptive filtering. A former data scientist said that you can even not use ML at all if you want.

Can someone really do DSP with these positions?

P.S. I know there are a lot of roles that have a better overlap with DSP (e.g. EEG, acoustics, radar, music information retrieval, ASR, media codecs) but I am focusing on the ones I mentioned, mostly because they seem to be doing a big impact today. And I believe someone would answer something like "it depends on how good you want to be at your job" but I'd like to consider roles where you are naturally forced to build DSP ad-hoc solutions.

$\endgroup$

closed as primarily opinion-based by Marcus Müller, Stanley Pawlukiewicz, MBaz, Peter K. Jun 3 at 12:48

Many good questions generate some degree of opinion based on expert experience, but answers to this question will tend to be almost entirely based on opinions, rather than facts, references, or specific expertise. If this question can be reworded to fit the rules in the help center, please edit the question.

  • 4
    $\begingroup$ Just one data point: I'm working on a problem where my old-fashioned hand-crafted solution is slightly outperformed by a multimillion node neural network that takes a week to train on a 250 watt GPU. The solution will be implemented on a battery-operated embedded processor -- no cloud-processing nonsense. Guess which solution will make it to the product? Now, if you're young and getting into the field, you'd do well to know at least the rudiments of ML. $\endgroup$ – MBaz Jun 2 at 16:08
  • $\begingroup$ It's nice to hear that! And I think that one can also argue that the ML solution doesn't offer the possibility to understand the knowledge stored in the layers (if a NN is used). Sure, ML knowledge is definitely needed, RNN and CNN are must-know tools nowadays, and there are more. In the end, DSP is still used in the convolutional layers of a CNN (to give an example). It's just that feature extraction is automated now, that's the problem: DSP is becoming less an art, and CV and DS roles are being performed by people with low mathematical background. $\endgroup$ – JFonseca Jun 2 at 21:43
  • $\begingroup$ Interesting question, but I believe the answer will definitely be primarily opinion-based... which the first line of Stan's answer suggests. $\endgroup$ – Peter K. Jun 3 at 12:49
  • 1
    $\begingroup$ Regarding "ad hoc" solutions: A friend of mine always used to say "Don't make a mockery of honest ad hockery". :-) $\endgroup$ – Peter K. Jun 3 at 12:50
4
$\begingroup$

in my opinion

your question seems based on a combination of thoughtful reflection and possibly, pre Alzheimer’s cognitive biases. ;)

There was an article in the WSJ not too long ago , I’m outside the paywall so I can’t provide a link, that looked at IBM’s efforts using Watson at major Cancer research centers.

The conclusion was that Watson wasn’t offering much in terms of originality. It wasn’t revealing those deep unifying subtle links in the masses of data it was being feed. It tended to find what was already known.

hard problems are still hard particularly if the way the data is represented ( labeled) isn’t relevant to what might be hidden. The threshold of what might be considered for brute force search might be advancing but NP hard hasn’t gone away.

I don’t believe I answered your question but I don’t think your question has a clear answer.

When I was an undergrad, the EE department had just the year before my junior semester had dropped land surveying as a core requirement. skill mixes change. not many years ago i was calculating viewsheds to estimate RF coverage. Land Surveying would have probably had some use.

Are patterns truth?

$\endgroup$
  • $\begingroup$ Thanks for your answer Mr. Pawlukiewicz and for the insights you provided on dsp.stackexchange.com/questions/52147/… (I think you deleted your answer in the past?) Yeah, Watson is an example of how deep learning can outperform tasks previously executed by expert systems (like Dendral) but it has its own limitation, no doubt about it. Yes, I also think that even modern AI solutions cannot solve all questions, particularly if they are NP-hard. So DSP would be still needed to reduce redundancy in high dimensional data (at least until quantum CPUs arrive). $\endgroup$ – JFonseca Jun 2 at 21:16

Not the answer you're looking for? Browse other questions tagged or ask your own question.