Abstract
Deep neural networks are notoriously susceptible to adversarial perturbations—small, human-imperceptible changes that lead to catastrophic misclassification. This paper introduces the concept of "computational entanglement" to explain how overparameterized models fit random noise and why these non-robust features persist. By mapping neural network layers to a physical spacetime framework, we demonstrate that maximizing model weights triggers a mathematical equivalent to relativistic length contraction. This allows models to "reconcile" seemingly unrelated features, effectively turning noise into structured information. Our findings suggest that adversarial vulnerability is a fundamental consequence of the geometric properties of high-dimensional weight spaces in deep architectures.
Introduction
A persistent challenge in machine learning is understanding why deep neural networks (DNNs) are so easily fooled by adversarial examples. While it is well-established that overparameterized models can memorize noise, a unified physical intuition for this behavior has remained elusive. Current theories often struggle to explain why "non-robust features"—patterns that are predictive but incomprehensible to humans—transfer so effectively across different model architectures.
This research bridges the gap between the abstract geometry of high-dimensional weight spaces and the practical reality of adversarial attacks. We propose that as data propagates through a deep network, it undergoes a process analogous to movement through physical spacetime. This perspective reframes adversarial examples not as simple errors, but as a form of "information reconciliation" where a network entangles a test sample with its internal representations to achieve a minimum energy (zero loss) state.
Methods: Spacetime Mapping and MaxLikelihood
The core of our theoretical framework involves treating the layer index ($L$) of a neural network as a temporal component ($ct$) and the feature dimensions as spatial components. By applying Pythagorean principles to this mapping, we derive equations for time dilation and length contraction within the network.
In this model, maximizing the weights of a network is mathematically equivalent to increasing the "velocity" of a feature vector. As this velocity approaches the theoretical "speed of light" within the network's manifold, the distance between disparate features contracts, causing them to converge toward a shared point.
To validate this theory, we introduced the MaxLikelihood algorithm. Unlike standard supervised learning, MaxLikelihood is an unsupervised optimization method that uses Cosine Distance-based Locality-Sensitive Hashing (LSH) to maximize the norm of the output vector at each layer. This forces the network to converge toward a specific normalized point in a flat Riemannian manifold, effectively "entangling" the input with a target state regardless of the input's initial structure.
Key Results: The Collapsing Loss Phenomenon
Our experiments using synthetic Gaussian noise and binary object images revealed a striking "collapsing" loss phenomenon. In overparameterized regimes (large $n$ and $L$), the loss distribution does not merely shift; it bifurcates and collapses symmetrically toward either 0 or 1.
This result demonstrates that the network is not simply learning a decision boundary in the traditional sense. Instead, it is "entangling" the test sample with the training data. We achieved zero loss for arbitrary test samples—even those never encountered during training—simply by increasing the number of layers and neurons.
Furthermore, we demonstrated the robustness of this entanglement by recovering a "Panda" image from a "Gibbon" reference point under extreme noise conditions. By treating the adversarial "Panda" as a noisy signal, a sufficiently deep network ($L=60$) could reconcile the hidden structure and recover the original data, confirming that adversarial examples contain enough latent structure for deep networks to exploit via relativistic convergence.
Discussion and Limitations
While our spacetime mapping provides a novel foundational perspective, it currently focuses on feedforward linear networks. Real-world deep learning relies on non-linear activations like ReLU, which may introduce additional complexities to the spacetime manifold. Additionally, while our empirical evidence is strong for binary classification and synthetic noise, further research is needed to extend these findings to multi-class, high-resolution datasets like ImageNet.
Ultimately, this work suggests that the "minimum solution" found via computational entanglement may be a primary driver of both the remarkable memorization capabilities and the inherent vulnerabilities of deep neural networks. Understanding these relativistic mechanics is a crucial step toward developing truly robust machine learning systems.
Summary updated