4
$\begingroup$

I intend to pursue a PhD program in digital signal processing (DSP). However, it seems nowadays that research has saturated in DSP and it's all about machine learning and deep learning. Are there are any areas in pure DSP that can still be explored for a PhD? If so, what are some research groups that focus on these areas?

$\endgroup$
1
  • 1
    $\begingroup$ Thank you for the validation $\endgroup$ Feb 14 at 19:58
6
$\begingroup$

Choosing a novel topic is a bold and wise move. Your question requires a bit of a bet, yet it is interesting to me, as I have not thought about this for a while. I will first drop two dual thoughts (they could be updated):

  • shed light on black-boxes: the present successes of machine learning and deep networks shall not make us forger that signals are our (DSP community) sensing door to a globally unknown world. Truly understanding what we can understand with data and models remains at our core, and we can be much better at that (to make it more interpretable, robust, model-based, explainable, reliable etc.)
  • save electrons in devices and computations: sensing and computing are becoming pervasive (IoT, etc.), and this requires materials (from the extraction of rare-earth elements for sensors and computing devices) and resources (form energy-harvesting sensing to bit-wise or qubit-related fast algorithms). Saving power (efficient filters, compression, fast algorithms, etc.) are in our DNA.

DSP is a blended field of knowledge, where people of mixed origins (computing, acquisition, information theory, physics, optimization, computer vision, statistics) meet. I hope that we DSP people can innovate here.

Additional stuff, as they come:

$\endgroup$
6
  • 1
    $\begingroup$ I completely agree with your point that we should "shed light on black-boxes". I have tried to look around for groups that explore interpretable machine learning, but unfortunately, there are very few and those that do are not taking on PhD students at this time. Thanks a lot for your answer. $\endgroup$
    – mhdadk
    Jan 23 at 11:34
  • 1
    $\begingroup$ There are many avenues to interpretability in machine learning and deep networks. Some rely on transforms (scattering networks) some on optimization (regularity conditions), some on geometry (manifolds), some on physics (ML & PDE), and many others. I am pretty sure that there are many open positions in those areas $\endgroup$ Jan 23 at 11:59
  • 1
    $\begingroup$ @mhdadk my PhD position happens to be pretty much exactly that: explainable / reliable machine learning (at least in how it was officially advertised). The way I got it was by having a background from physics plus experience in CS (this is an informatics institute), that convinced them. If you're applying for a similar position, strong DSP foundations are probably a similarly good selling point. This may also work if you're applying for an ML position that's not specifically focused on interpretability or so – after all, a PhD is supposed to involve own ideas on the research direction. $\endgroup$ Jan 23 at 15:21
  • 2
    $\begingroup$ (Just note that to actually publish in the AI field, you won't get far with mostly theoretical DSP arguments... all that really convices reviewers there is results on real-world data.) $\endgroup$ Jan 23 at 15:21
  • $\begingroup$ @LaurentDuval do you know of any papers/references that discuss these different approaches to interpretability? $\endgroup$
    – mhdadk
    Jan 27 at 11:52

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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