design encoder-decoder models

Conventional decoders for dense prediction tasks suffer from outdated architectural designs and insufficient cross-layer contextual sharing, limiting feature propagation efficiency and spatial consistency. Method: We propose a novel decoder architecture centered on a learnable, shared “bank”—a parameterized module dynamically resampled and fused across multiple scales to enable explicit cross-layer contextual reuse during decoding, thereby departing from traditional serial, layer-wise independent decoding paradigms. Built upon a Transformer backbone, the bank is jointly optimized end-to-end. Contribution/Results: Our approach significantly improves decoding efficiency and spatial coherence. On both natural and synthetic image depth estimation benchmarks, it substantially outperforms state-of-the-art methods, achieving superior accuracy and generalization under large-scale training. To our knowledge, this work presents the first systematic design and empirical validation of a universal, decoder-level contextual sharing mechanism.

This work clarifies the functional boundaries and necessity of core Transformer components—tokenization, embedding/un-embedding, masking, positional encoding, and padding—addressing widespread conceptual ambiguity in their mechanistic roles. Targeting ML engineers, we propose an incremental, invertibility-based analytical framework: using binary (0/1) sequences as probes, we systematically introduce each component via a “zero-one construction” and empirically validate its irreplaceability in the encode-decode pipeline. Implemented lightweightly in PyTorch, our framework supports manual attention matrix construction, explicit positional embedding injection, and interpretable mask design. Experiments demonstrate significantly improved conceptual accuracy among learners. Notably, we provide the first empirical verification that, in the absence of self-attention, positional encoding combined with padding alone suffices for basic length-aware tasks.

This paper systematically surveys the evolution of neural network design paradigms in computer vision, focusing on image recognition, generative modeling, and self-supervised learning. Through a comprehensive literature review and cross-model comparative analysis, we establish a unified analytical framework that identifies three overarching trends: (i) convolutional architectures giving way to attention-based mechanisms; (ii) supervised learning transitioning toward self-supervised paradigms; and (iii) discriminative modeling shifting to generative modeling. We conduct in-depth analyses of six landmark models—ResNet, ViT, GAN, LDM, DINO, and MAE—highlighting their breakthroughs in training stability at depth, generative fidelity, and reduced reliance on labeled data. Crucially, we unify diverse mechanisms—including residual connections, momentum teachers, masked encoding, diffusion processes, and adversarial training—within a coherent design logic, revealing intrinsic consistency across architectural advances. Our synthesis yields reusable architectural principles and theoretical guidance for next-generation vision foundation models. (149 words)

To address excessive computational overhead when deploying vision models on resource-constrained devices, this paper proposes an efficient Vision Transformer (ViT) architecture design framework. Methodologically: (1) it optimizes the input-output data pathway to enhance representational capacity of lightweight models; (2) it restructures the context window of computationally constrained attention mechanisms to improve local-global modeling efficiency; and (3) it leverages the invertibility and explicit probabilistic modeling properties of normalizing flows to enable high-fidelity, low-overhead knowledge distillation. Experiments demonstrate that the proposed approach achieves comparable or superior accuracy on benchmarks such as ImageNet, while requiring significantly fewer parameters and FLOPs. It also substantially reduces inference latency and memory footprint. The framework establishes a scalable new paradigm for efficient visual understanding at the edge.

To address high first-token latency and low throughput of small models (≤1B) on edge devices, this work revisits the encoder-decoder architecture for its efficiency advantages under resource constraints. We propose a task-adaptive knowledge distillation framework that enables lightweight encoder-decoder student models to effectively inherit capabilities from large decoder-only teachers while preserving their intrinsic properties—single-pass input encoding and decoupled understanding and generation. This is the first systematic validation of such architectures on asymmetric-sequence tasks. Integrated with RoPE, cross-platform optimization (GPU/CPU/NPU), and fused visual encoders, our approach achieves a 47% reduction in first-token latency, a 4.7× throughput improvement, and an average +6-point gain in task-specific performance—particularly pronounced in tasks with large input-output distribution mismatch.

This work addresses the growing complexity and lack of interpretability in deep image compression autoencoder models, which hinder the design of efficient architectures. For the first time, it systematically employs Jacobian analysis to examine the internal transformations of unbiased autoencoders, uncovering consistent and interpretable operational patterns that are prevalent across high-dimensional compression models. Building on these insights, the study identifies multiple semantically meaningful internal operations shared across diverse models and demonstrates their utility in constructing lightweight architectures that simultaneously achieve high compression performance and low computational complexity. This approach establishes a new paradigm for designing interpretable and efficient compression models grounded in analytically derived internal mechanisms.

This study decodes visual information from high-density neural recordings in the primate cortex to investigate how neural activity underpins perception. We systematically evaluate the impact of model architecture, training objectives, and data scale on decoding performance, proposing an efficient decoder that combines a lightweight temporal attention module with a shallow multilayer perceptron. Furthermore, we introduce a generative framework integrating low-resolution image reconstruction with semantic-conditioned diffusion. Experiments demonstrate that our approach achieves 70% Top-1 accuracy on image retrieval tasks, substantially outperforming existing methods. Our findings also reveal diminishing returns with increasing input dimensionality and dataset size, underscoring the critical role of temporal dynamics modeling in visual neural decoding.

This work addresses the challenge that generative 3D models often fail to satisfy additive manufacturing constraints—such as overhang angle, minimum wall thickness, and structural strength—by proposing a neural decoder–based deep learning framework that enables, for the first time, end-to-end generation of printable 3D geometries directly from latent representations. The method explicitly embeds multiple manufacturability constraints into the decoder training process, jointly optimizing geometric validity and printability. Experimental results demonstrate that the generated structures exhibit high manufacturability across diverse object categories and have been successfully validated through physical 3D printing, significantly outperforming existing generative approaches in both feasibility and fidelity.