Modern computer vision from CNNs to foundation models: a unified framework for representation learning
Abstract
Modern computer vision is increasingly driven by the search for representations that generalize across tasks, modalities, and learning settings, rather than excelling at a single benchmark. This has produced a broad collection of architectures and paradigms, including convolutional neural networks (CNNs), Vision Transformers (ViTs), state-space models, foundation models, and diffusion-based generative models. Despite differing in design and training objectives, these approaches increasingly draw on a common set of principles: representation learning, self-attention, self-supervision, multimodal alignment, and generative modeling. As a result, the boundaries between discriminative, multimodal, and generative paradigms are becoming progressively less distinct. This survey examines that convergence through the lens of representation learning. We introduce a formal framework that classifies architectures by which component of an encoder–decoder pipeline they modify, unifies diverse training objectives as instances of a single alignment functional, and formalizes the locality–globality trade-off through two provable relationships linking a model’s locality, its computational cost, and the depth required to reach a global receptive field. Using this framework, we trace the evolution of vision architectures, from convolutional networks and Vision Transformers through hybrid and state-space designs to multimodal foundation models and modern generative frameworks, critically examining the limitations of each paradigm rather than reporting only their successes and complementing this analysis with a quantitative comparison across representative architectures. We show how discriminative, multimodal, and generative learning are converging around shared representational principles, argue for a unified representation-learning perspective on modern computer vision, and outline open challenges and concrete future directions toward general-purpose visual intelligence, multimodal foundation models, and trustworthy deployment.