Convolutional encoders are widely used in modern artificial intelligence systems to transform structured inputs into compact representations that are subsequently processed by pooling, flattening, and training-based classification layers. Despite their empirical success, this pipeline implicitly assumes that learning is intrinsic to convolutional processing. In this work, we show that convolution itself is a deterministic linear measurement operation and does not inherently require training; learning becomes necessary only after architectural choices discard geometric structure and invertibility. By reformulating convolutional encoding as a known forward operator, inference is cast as an inverse problem governed by algebraic consistency rather than optimization trajectories. When spatial structure is preserved and pooling and flattening are avoided, the encoded representation admits a σ-regularized equilibrium solution obtained via the adjoint convolution operator. This formulation yields a unique closed-form reconstruction in a single computational step, eliminating gradient descent, backpropagation, learning rates, and iterative updates, and resulting in deterministic, reproducib
📖 افتح في inklap 🔗 DOI 📮 اطلب بحثاً