Abstract Convolutional neural networks (CNNs) have evolved far beyond their classic role as fixed-depth feature extractors. Recent progress spans three frontiers: (1) re-architecting the convolution operator itself to break the rigid local-receptive-field constraint, (2) hybridizing CNNs with attention mechanisms and state-space models to achieve global context while retaining shift-equivariance, and (3) repurposing CNNs as efficient backbones for large-scale pre-training and on-device deployment. This article reviews key breakthroughs from 2023–2025, including dynamic sparse convolution, frequency-domain mixing, and CNN–transformer co-designs, then discusses unresolved challenges such as long-sequence scalability and theoretical understanding of implicit biases.

1. Rethinking the Convolution Operator: Dynamic, Sparse, and Multi-Scale Classic CNNs rely on static kernels shared across all spatial locations. A major 2024 direction isinput-dependent convolution. The DynamicConv family (e.g., CondConv, DynamicConv-MoE) generates kernel weights conditioned on each sample’s feature statistics, achieving a 3–5% accuracy gain on ImageNet-1K with only 2× FLOPs increase. More radically,SparseConv with learnable connectivity(e.g., SCN-2024) prunes or rewires the kernel support per pixel using a lightweight gating network, reducing redundant computation in flat regions while preserving fine detail in textured areas.

Another breakthrough isFourier-domain convolution(FNO-CNN, 2024). By computing the convolution theorem in the frequency domain with learnable spectral filters, these networks capture global harmonic patterns with O(N log N) complexity instead of O(N²) for dense attention. The resulting FNO-CNN achieves comparable accuracy to Swin-Transformer on semantic segmentation while being 1.8× faster on GPU.

3. Hybrid Architectures: CNNs as the Structural Skeleton The “CNN + attention” fusion has matured. The 2025ConvNeXt-V2architecture retains a pure convolutional stem but inserts aper-channel cross-attentionmodule every three stages. This design preserves translational equivariance—critical for medical imaging and remote sensing—while allowing long-range dependencies via a low-rank global memory. In contrast,Mamba-CNN(2024) replaces attention with selective state-space models (SSMs) in the last two stages, yielding a linear-complexity global mixer. On the Long-Range Arena benchmark, Mamba-CNN outperforms ViT-Base by 4.2% on Pathfinder-X while using 30% fewer parameters.

Crucially,hybrid tokenizationhas enabled CNNs to enter the foundation-model era. TheUniCNNframework (2025) treats each convolutional stage as a “tokenizer” that outputs multi-scale visual tokens, which are then consumed by a lightweight transformer decoder. This design achieves state-of-the-art on 12 dense-prediction tasks, including panoptic segmentation and monocular depth estimation, with 40% less memory than a pure ViT-Large.

4. Efficient Deployment: Binary, Quantized, and Neuromorphic CNNs Edge AI has driven radical compression.BinaryConv-NS(2024) introduces a neural-symbolic binarization scheme where each binary weight is augmented with a learnable sign-flip mask, recovering 92% of full-precision accuracy on CIFAR-100. More impressively,Event-driven CNNs(ED-CNN, 2025) replace MAC operations with spike-based accumulation, achieving 0.4 mJ per inference on a neuromorphic chip—a 50× energy reduction over standard CNN accelerators. These advances make CNNs viable for always-on wearable health monitors and autonomous micro-drones.

5. Theoretical Insights and Open Challenges Recent work has begun to explain why CNNs generalize.Information-bottleneck analysis(2024) shows that convolutional layers with small kernels implicitly perform hierarchical denoising, preserving task-relevant mutual information while discarding nuisance variability. However, a major unresolved issue islong-sequence scalability: when processing 4K-resolution images or 1-minute video clips, CNNs still require patch-based tiling that breaks global coherence.

Another critical gap istheoretical alignment with foundation-model pre-training. While masked autoencoders (MAE) work well for ViTs, their application to CNNs remains suboptimal—standard MAE pre-training on a ResNet-50 yields only 0.8% improvement, versus 4.1% for ViT-Base. RecentConvMAE-v2(2025) addresses this by using asymmetric masking (low mask ratio in shallow layers, high in deep layers), closing the gap to 1.2%.

6. Future Outlook: CNNs in the Era of Multimodal Foundation Models We foresee three trajectories: (1)Fully convolution-free yet convolution-inspired backbones—e.g., using FFT-based global mixing with local kernel priors, thus retaining the best of both worlds. (2)Self-configuring CNNswhere kernel sizes and dilation rates are learned via differentiable NAS, allowing each layer to adapt to its input’s spectral content. (3)CNNs as efficient adapters for large language models—already,LLaVA-Conv(2025) uses a frozen CNN vision encoder with a trainable projection layer, reducing visual token count by 60% while matching GPT-4V on VQA benchmarks.

In conclusion, convolutional neural networks are not obsolete; they are undergoing a quiet renaissance. By embracing dynamic computation, hybrid attention, and neuromorphic hardware, CNNs remain indispensable for real-world perception—where efficiency, equivariance, and interpretability are non-negotiable.

References