Abstract: Deep learning has made an enormous impact on computer vision in the last decade. What ML technologies lie ahead of us that may have an impact on the future of computer vision? I will discuss two such general technologies: equivariance and quantum statistics. Equivariance plays an important role in making our models of images independent of some arbitrary choices that we are forced to make in a computer implementation, such as the choice of the pixel grid. Farther away in the future I also predict that quantum computation could play an important role in processing of image and video data. I will try to introduce quantum mechanics at a very simple level and hope to make a convincing argument why optical quantum computers are in fact a very natural match for computer vision problems.
Beyond Brute-Force Learning
Over the past decade, deep learning has fundamentally reshaped the landscape of computer vision. However, as we look toward the next ten years, Dr. Max Welling suggests that the current paradigm—characterized by "brute-force" learning on classical hardware—may be reaching its limits. To advance, the field must move beyond models tied to arbitrary digital artifacts and toward architectures that respect the underlying physics of the world. Welling identifies two critical technologies for this transition: equivariance and quantum statistics.
Breaking the Tyranny of the Pixel Grid
A central critique in Welling’s vision is our reliance on the pixel grid. While pixels are a convenient engineering standard for digital displays and sensors, they are not a fundamental property of the physical world or biological vision. Current models are often sensitive to the specific alignment, resolution, or sampling of these grids.
This "tyranny of the pixel grid" forces models to learn representations that are artifacts of the computer rather than the scene itself. To reach the next level of abstraction and efficiency, computer vision needs to become more "physics-aware" in how it handles geometric symmetries.
Pillar 1: Equivariance as a Mathematical Foundation
To address the limitations of grid-based learning, Welling proposes a shift toward equivariance. Equivariance is a mathematical framework ensuring that if an input undergoes a transformation—such as a rotation or translation—the model’s internal representations change in a predictable, corresponding way.
By "baking in" these symmetries as a first principle, we can treat images as continuous signals rather than discrete blocks of data. This approach allows for models that are more robust and independent of the arbitrary choices made during digital implementation, such as the specific orientation of a pixel array. Instead of learning every possible rotation of an object from scratch, an equivariant network understands the relationship between these views inherently.
Pillar 2: The Quantum Future of Vision
Looking further ahead, Welling argues that quantum computation—specifically optical quantum computing—represents the next frontier. This is not merely a quest for faster processing; it is a search for a more "natural match" for the problem of vision.
Vision is fundamentally the study of light interacting with the environment. Since light is governed by the laws of quantum mechanics, processing visual data using quantum statistics is a more direct mapping of physical reality than translating it into classical binary logic. Unlike standard silicon-based quantum bits (qubits), optical quantum systems use photons—the very medium of vision—to perform calculations.
While Welling acknowledges that quantum computation for computer vision is "farther away in the future," the theoretical alignment is profound. Moving from classical logic to quantum statistics could allow us to process image and video data in a way that aligns with the fundamental nature of light itself.
Conclusion
The next decade of computer vision will likely be defined by a move away from engineering conveniences and toward physical principles. While the integration of quantum mechanics remains a long-term goal, the move toward equivariance is already providing a roadmap for more efficient, robust models. By focusing on symmetry and the physical nature of light, the research community can transition from brute-force pattern matching to a more principled, physics-aware understanding of the visual world.
Summary updated