Finding Signals in the Noise
From hidden patterns to usable systems
Much of science is the art of finding signal in noise. Looking back, that idea connects every stage of this journey, from algebra and signal processing to neural networks, computer vision, and speech technologies.
Looking back, the journey was never driven by a grand plan. Rather, it unfolded through a series of small discoveries, each opening a door to a larger world.
One of the earliest lessons was that aptitude is often acquired rather than inherited. As a schoolboy, mathematics did not come naturally. It was only around Class V or VI that algebra began to make sense, thanks to my father. What had first appeared to be a collection of mysterious symbols revealed an underlying logic. There was a quiet satisfaction in realizing that abstract notation could describe concrete relationships. In retrospect, this may have been the first encounter with a recurring theme: hidden structure waiting to be uncovered.
Access to computers was scarce during those years. Even later, during undergraduate studies at College of Engineering, Jawaharlal Nehru Technological University Kakinada, computing resources were very limited. In hindsight it was a boon, allowed one to concentrate on the theory. During my Masters at Regional Engineering College (REC), Tiruchirappalli, most student work was conducted through local area network (LAN) terminals connected to central systems rather than on personal machines. Computing was precious, measured and often indirect. Every minute of access mattered, and experimentation required patience.
The fascination with signals emerged much later. During the final semester of engineering, a course on digital signal processing provided an unexpected revelation. Initially, it was the long equations filling the blackboard, becoming even longer when extended to two-dimensional signals, that captured my attention. Signals were not merely waveforms plotted in textbooks; they were mathematical objects that could be transformed, analyzed, filtered, and understood. Concepts such as convolution, Fourier analysis, and spectral representations offered a way of looking beneath the surface of observed data. For perhaps the first time, mathematics, computation, and real-world phenomena converged into a coherent discipline.
Fortunately, this interest carried forward into postgraduate studies in satellite communications. Master's level courses provided a deeper exposure to signal processing and strengthened the mathematical foundation required to appreciate the subject. During this period came an introduction to artificial neural networks, which, in the late 1980s and early 1990s, remained largely an academic curiosity rather than the dominant force they would later become
Unlike today, there were no readily available software libraries, frameworks, or examples to build upon. Understanding a neural network meant understanding every equation and every line of implementation. Fascinated by the idea that machines could learn from examples, a simple feedforward neural network was implemented entirely in C. The architecture was modest: two input nodes, a single hidden layer, and one output node. Training was performed using the back-propagation algorithm. The objective was equally simple but intellectually captivating: solving the XOR problem, a classic example that demonstrated the limitations of the single-layer perceptron and the power of hidden representations. Looking back, unknowingly, I was also doing data augmentation because the XOR problem had just four data points and to train one required a large number of training data.
The success of that implementation left a lasting impression. The notion that a machine could discover decision boundaries through iterative adjustment of weights seemed remarkable. More importantly, it suggested possibilities far beyond toy examples This conviction led to an opportunity during a sponsored project with the Naval Physical and Oceanographic Laboratory (NPOL). The task involved passive sonar target classification, an area traditionally approached through building an expert system with handcrafted features. Neural networks were proposed as a possible solution mechanism for classification using hand-crafted features. At the time, this was far from a mainstream choice. The resulting work led to a first publication, providing an early glimpse of how learning-based techniques could complement conventional expert systems.
In parallel, during my masters an image processing system (Matrox frame grabber boards installed on a PC) was procured which allowed practical implementation of image processing theory. This led to a deeper engagement with image processing through the development of a functional OMR (Optical Mark Recognition) sheet correction system Exposure to neural network and image processing fueled a desire to pursue deeper research in computational approaches to signals, images and pattern recognition.
That enthusiasm eventually led to IIT Bombay, where the intention was to pursue research centered on neural computation as applied to image processing and analysis. However, the intellectual journey that followed was not one of continuous affirmation. During this period came a closer encounter with the influential critique found in Perceptrons by Marvin Minsky and Seymour Papert. Reading it produced a profound sense of disillusionment. The limitations highlighted in the book cast doubt on the future of neural-network research, particularly in an era when computational power remained severely constrained and the availability of data for training neural networks was equally limited.
As a result, attention gradually shifted elsewhere.
Conventional image processing and computer vision emerged as appealing alternatives. They offered rich mathematical structure, challenging computational problems, and tangible visual results. Yet even these pursuits were shaped by the limitations of contemporary hardware. Research often involved working on Intel 386-based systems whose hard disk capacities were smaller than today's random access memory (RAM) and computational resources that would appear almost unimaginable by today's standards,
Simple matrix operations could become significant undertakings. Computing something as small as a 16x16 matrix multiplication required careful attention to implementation and efficiency in terms of freeing the buffer before assigning it to another varlable. Data movement itself was frequently part of the workflow. Experiments were conducted through a ritual familiar to many researchers of that era: inserting floppy disks, transferring files, running a computation, waiting, and repeating the cycle. Progress was measured not only in ideas but also in patience.
The doctoral years brought a return to signal processing, now at a deeper and more mathematical level. A course on wavelet transforms expanded the conceptual toolkit beyond traditional Fourier methods. Wavelets offered a framework for understanding signals simultaneously in time and frequency, enabling representations that were localized, adaptive, and intuitive. They reinforced a growing appreciation that no single representation is universally optimal; rather, insight often emerges from viewing the same phenomenon through multiple lenses.
The postdoctoral period added another important layer to this journey by revealing the practical use of theory, especially through work on stereo vision. Earlier, image processing often began with a given image: one accepted the image as input and then applied enhancement, segmentation, recognition, or measurement techniques to it. The camera that produced the image was seldom questioned. In stereo vision, however, the image could no longer be treated as an isolated object. Working in the three-dimensional world based on what cameras observed required understanding how the physical world was projected onto a two-dimensional image plane, how camera parameters shaped what was recorded, and why calibration was essential for reliable reconstruction. Concepts such as perspective projection, camera geometry, correspondence, and depth estimation turned mathematical theory into working practice. This experience opened a deeper appreciation of theory not as abstraction alone, but as something that must explain, guide, and withstand the imperfections of real devices and real measurements.
These learnings continued to influence later work in ways that were not always obvious at the time. The understanding of cameras, projection, geometry, and reconstruction naturally fed into attempts to build virtual try-on systems for eyewear, jewellery, and clothing well before such applications became widely discussed or commercially visible. These efforts required the same movement from abstract models to usable systems: an image was not just an image, but a measurement shaped by viewpoint, scale, ilumination, calibration, and the geometry of the object and the human body. The postdoctoral experience also left a more durable lesson that carried into speech research: technology becomes meaningful only when it can be built into systems that people can actually use Whether the input was an image, a three-dimensional scene, or a speech signal, the challenge remained the same, recovering structure from imperfect observations and turning that structure into practical tools.
In hindsight, the trajectory appears less fragmented than it felt while living through it. Algebra, signal processing, neural networks, image processing, computer vision, wavelets, and postdoctoral work in stereo vision all revolved around a common objective: discovering structure in complex data and making that structure usable. Each stage offered a different perspective on the same challenge. Algebra revealed hidden relationships behind symbols; signal processing and wavelets showed how structure could be exposed through suitable representations; neural networks suggested that patterns could be learned from examples; image processing and computer vision brought computation closer to visible phenomena; and stereo vision demonstrated that theory acquires its full force only when it engages with cameras, calibration, physical geometry, and imperfect measurements. That lesson stayed on in later attempts to build practical systems, from early explorations of virtual try-on for eyewear, jewellery, and clothing to subsequent work in speech research. Across these efforts, the goal was not merely to analyze data, but to turn analysis intc systems that could interpret the world, support decisions, and be of practical use.
What began as an effort to understand a simple algebraic equation eventually evolved into a lifelong fascination with extracting meaning from signals, images, and patterns. The technologies changed the computational resources expanded dramatically, and entire fields rose, faded, and re-emerged. Yet the underlying motivation remained remarkably constant the desire to understand how information is represented, transformed, and ultimately interpreted
In that sense, the journey was never really about moving from one field to another. It was about following the signal wherever it happened to lead.
Comments