Contents The year 2023 marked the 75th anniversary of the IEEE Signal Processing Society (SPS), which was founded in 1948 as the “Professional Group on Audio” of the Institute of Radio Engineers (IRE), becoming the first IEEE Society. (The IRE, founded in 1912 with a focus on radio and then electronics, together with the American Institute of Electrical Engineers, founded in 1884 with an emphasis on power and utilities, were united in 1963 to form IEEE.) At ICASSP 2023, in Rhodes, Greece, I chaired a panel with Alex Acero of Apple, K. J. Ray Liu of Origin Wireless, Ali H. Sayed of EPFL, and Rabab K. Ward of the University of British Columbia that discussed the future of signal processing (SP) with a focus on the relations between artificial intelligence (AI) and machine learning (ML) and SP. Ours followed a panel chaired by Mos Kaveh, University of Minnesota, with Anthony Constantinides, Imperial College London, Alan Oppenheim, Massachusetts Institute of Technology, and Ron Schafer, Georgia Tech, as panelists who reminisced on the origins of SP.
Our panel tackled the elephant in the room to which we came back repeatedly. When AI is taking the public imaginary and collective consciousness by storm, should we just accept becoming an “AI subfield” or, on the contrary, taking a hint from the past, recognize that we are more relevant and dynamic than ever?
Let me argue the latter and remind ourselves and the world of our unique value and significance.
First, let me state that I take a broad view of what SP has stood and stands for. Where there are data and a willingness to make some intelligent statement and take some meaningful action based on data, there is SP. Processing signals and data have been around possibly since the dawn of humankind, and SP has roots that extend well before 1948, back to Fourier, Gauss, or much earlier. We usually confuse the founding of SP as a discipline, along with its formal methods taught in academia and used by technologists, with the founding of the SPS. From these early beginnings to this day, SP is pervasive, underlying many of the advances in our digital world. To be clearer, we need to look only to the latest few issues of this magazine. Commemorating the 75th-year event, IEEE Signal Processing Magazine published in the June and July 2023 issues several articles on the major accomplishments in the field in the last 25 years, covering advances in many diverse areas, from multiresolution, wavelets, and sparse SP, to audio and speech processing, brain–computer interfaces, multimedia SP, information forensics, radar, communications, super-resolution imaging, and the infusion of optimization in many traditional SP areas, like beamforming and array processing, to the emergence of new areas, such as computational imaging, distributed and decentralized networked inference, and graph SP. These articles are well worth reading and show that our current state of affairs reflects a highly dynamic field with major impactful contributions. Even if we paraphrase an Intel market gimmick of the 1990s and 2000s, “SP inside,” we SP’sers know for a fact that SP has been a major contributor to the digital wireless and AI revolutions.
Now, back to the future. With this background, it was fitting that we peeked into what lies ahead of SP as a discipline and as a community, starting with the present and looking forward to 25 years from now. The present is dominated by three major factors: 1) data of all forms—physical, social, corporate, biological, chemical, and genetic—in large (big) amounts, everywhere collected, produced, stored, and processed; 2) computing, available everywhere and anywhere, fast, with ever-evolving architectures, continuing to beat Moore’s law; and 3) algorithms, including the fast NlogN algorithms, like the fast Fourier transform, augmented now with deep learning models and convolutional and graph neural networks. The confluence of all of these factors has led to a sudden explosion of new possibilities, with yesteryear’s models of few parameters replaced by models that count billions if not trillions of parameters, and a public perception of an AI-and-ML-dominating world. The commanding perception in many communities is that any decision making and actions inferred from data are AI, and that it is practiced by some form of deep learning model. Since SP is inference, learning, adaptation (remember the old equalizers from the 1960s?), and extraction of actionable information from reams of data, one might conjure (with a rather narrow view) that SP is now rebranded as AI. This is the elephant in the room. While we acknowledge some clear correspondences between these two domains, it is important to recognize that the discipline of SP admits a broader perspective: it deals with both models and data, and it pays particular attention to limits of performance, theoretical guarantees, and the engineering and robustness of deployed systems. In the past, we would distinguish “gray boxes,” data-driven and adapted models, from “black boxes.”
Taking a clue from the past, SP has been most successful when developing new theoretical methods that shed new light and understanding on significant problems and that lead to solutions that work in the real world. The real world is noisy, departs from abstractions and models, is limited by finite resources, and entails compromises and tradeoffs. SP excels at this; it recognizes the prevalence of data, but it does not ignore the prior knowledge or any domain knowledge and expertise that are available. SP builds systems that, once deployed, must be robust, explainable, trusted, have performance guarantees, and last in many different environments, often using only scarce and limited resources.
Some successful examples illustrate these points. We may say that, in the early 1960s, speech research had a strong presence within the acoustics community, dominantly housed then in the Acoustical Society of America. These early studies relied on good models and statistical prediction methods, but speech researchers also felt the need for extensive evaluation with speech recordings, moving the field toward a more data-engaged community of speech engineers—namely, the SPS community. Image and video compression is another prominent example: the basic and fundamental methodologies underlying the multiple suites of codec standards released since the 1990s, starting with JPEG in 1992, benefited from the vigorous research and testing carried out by SP researchers. Many of the results were published in SPS transactions and conferences, eventually leading to the launch of the then-new IEEE Transactions on Image Processing and the first ICIP.
Video coding and many other technologies to which SP significantly contributed, including wireless communications, have given rise to many new interactive ways of communication, which we all experience today. In addition, new SP methods to analyze data have led to advances and the explosion of new application domains in the health sciences, like “medical and biological imaging, digital pathology, molecular imaging, microscopy, and associated computational imaging, image analysis, and image-guided treatment, alongside physiological signal processing, computational biology, and bioinformatics” [2]. Take areas like array processing, more traditionally applied to radar and sonar. “[C]ast[ing] [them] as optimization problems with side constraints originating from prior knowledge regarding the structure of the measurement system” [5] has led to a burst of new model- and data-driven applications with “impact in everyday life, including beamforming for ultrasound imaging, synthetic aperture radar for remote sensing, vehicular radar (ultrasound and electromagnetic) for autonomous driving, microphone arrays for human-machine interfaces (a good example is the Amazon Echo), and MIMO antenna arrays for Wi-Fi and mobile communications standards (IEEE 802.11n, IEEE 802.11ac, 3 G, WiMax, and LTE)” [4]. As our world turned increasingly digital, data collection became widespread and used methods involving multiple modalities. This led to the emergence of new distributed and decentralized ways of processing data, as described in [8], and new methods of processing data, indexed generically by agents, beyond time and pixel samples [3]. These are some examples in which algorithms + data + prior knowledge + extensive testing have broadened the impact of SP research in the real world. They illustrate how SP and the SP research community kept growing in terms of both new research and methods as well as new application domains with real impact in the world, growing and attracting new devoted young researchers to the field.
Looking ahead, beyond traditional domains, the world faces new challenges, in fact, “grand challenges,” presenting new opportunities to SP and the SP community. The new terms defining the new world and the new reality include, among many others, sustainability, cybersecurity, climate change, green energy, the safety of critical structures, and health. These topics are highly interdisciplinary, calling for new models, new algorithms, and new ways to process data. Some may label this as AI or ML, but we know that the SP way of thinking is at its heart. To build further success, we need to engage and tackle these applications using all of the tools we have available—including, of course, optimization and deep models—developing fair and explainable solutions that are robust and efficient. This leads me to a final thought: SP is an ever-expanding collection of methods and algorithms. From my perspective, SP has no qualms about broadening itself, when it makes sense. It started from two different threads: a digital SP focus (think Fourier and a book like Discrete-Time Signal Processing [6]) and a more noise-driven direction (think Wiener and a book like Detection, Estimation, and Modulation Theory, Part I [7]). These two directions fully converged in the 1980s in our current understanding of SP, where signals and data may be deterministic or random, and methods may be transform or statistically based. Over the years, the SP portfolio kept broadening as new tools and methods were developed in ours and many other related disciplines and were successful in processing data, as wavelets, compressed sensing, and matrix/tensor factorizations [1], or the widespread adoption of optimization and gradient descent illustrate. So, now, given their success, we enlarge our portfolio to include deep models and graph neural networks.
SP as a discipline is alive and kicking, as demonstrated by the many thousands of paper submissions to our journals, conferences, and workshops. SP has its own culture; it engineers solutions for relevant problems. We develop theories that validate our methods and provide performance guarantees. We deploy our solutions to work in the real world—solutions that are well grounded in theory and analysis and that are robust and work well under a variety of noise and disturbance conditions. SP is theory but also data, and that is its key strength. SP’s success will always be measured by the impact of our solutions on our lives and society at large.
For me “the elephant” is becoming increasingly small; after all, I am well used to SP as a stealth technology. Is SP a “subfield of AI?” Not for me, but I will let you decide.
ACKNOWLEDGMENT
I thank my copanelists, Acero, Liu, Sayed, and Ward, for sharing their perspectives and for their help with this column. I thank Tulay Adali, editor-in-chief of this magazine, for her invitation as well as her encouragement and feedback.