Native Vision Transformer (ViT) Overview

• Native Vision Transformer (ViT) is a pure transformer-based model that tokenizes images into fixed-size patches to perform image recognition without convolutional inductive bias.
• It processes patch embeddings with a standard transformer encoder architecture featuring multi-head self-attention, Pre-LayerNorm, and residual connections to capture both global and local features.
• The model achieves state-of-the-art results on benchmarks by leveraging large-scale pre-training and scaling depth, width, and patch size for optimal accuracy and computational efficiency.

The Native Vision Transformer (ViT) is a pure Transformer encoder model, originally devised for natural language processing, adapted for image recognition tasks without convolutional architectures. By tokenizing images into a sequence of fixed-size non-overlapping patches and processing them directly with the standard Transformer encoder architecture, ViT demonstrates that explicit convolutional inductive bias is unnecessary for high-accuracy image classification, provided large-scale pre-training. When pre-trained on extensive datasets such as ImageNet-21k or JFT-300M and transferred to benchmarks including ImageNet, CIFAR-100, and VTAB, ViT attains state-of-the-art results with reduced computational training costs relative to high-capacity convolutional neural networks (CNNs) (Dosovitskiy et al., 2020).

1. Patch Embedding and Input Representation

ViT operates by partitioning an input image $x\in\mathbb{R}^{H\times W\times C}$ into a regular grid of non-overlapping patches of size $P \times P$, yielding

$N = \frac{H\times W}{P^2}$

distinct patches. Each patch, $\mathrm{patch}_i(x)\in\mathbb{R}^{P^2C}$, is linearly projected via $E \in \mathbb{R}^{(P^2C)\times D}$ into a $D$-dimensional embedding:

$$x_p^i = E \cdot \mathrm{patch}_i(x) \in \mathbb{R}^D.$$

A learnable classification token $[{\tt CLS}]\in\mathbb{R}^D$ is prepended, and learnable positional encodings $E_{\text{pos}} \in \mathbb{R}^{(N+1)\times D}$ are added, forming the input sequence:

$$z_0 = \left[\,[{\tt CLS}];\,x_p^1;\,\dots;\,x_p^N\,\right] + E_\mathrm{pos}.$$

2. Transformer Encoder Architecture

The Transformer encoder processes the $P \times P$0-length sequence through $P \times P$1 identical layers, each comprising:

• Pre-LayerNorm
• Multi-Head Self-Attention (MSA) with $P \times P$2 heads:

$P \times P$3

where

$P \times P$4

and $P \times P$5, $P \times P$6, $P \times P$7, with $P \times P$8.

• Residual connection
• Pre-LayerNorm
• Two-layer MLP:

$P \times P$9

• Residual connection

The stack updates as

$N = \frac{H\times W}{P^2}$0

After $N = \frac{H\times W}{P^2}$1 layers, the output associated with the $N = \frac{H\times W}{P^2}$2 token is used for image-level classification, following a final LayerNorm and linear classification head.

3. Model Variants and Scaling Dimensions

ViT explores the scaling of depth, width, and patch size. The canonical configurations, using the “ViT-X/$N = \frac{H\times W}{P^2}$3” notation ($N = \frac{H\times W}{P^2}$4 for Base, Large, Huge; $N = \frac{H\times W}{P^2}$5 is patch size), are summarized below.

Model	Layers ($N = \frac{H\times W}{P^2}$6)	Dim ($N = \frac{H\times W}{P^2}$7)	Heads ($N = \frac{H\times W}{P^2}$8)	Params	Patch Size ($N = \frac{H\times W}{P^2}$9)
ViT-B/16	12	768	12	86M	16
ViT-L/16	24	1024	16	307M	16
ViT-H/14	32	1280	16	632M	14

Smaller patches (lower $\mathrm{patch}_i(x)\in\mathbb{R}^{P^2C}$0) lead to longer token sequences ($\mathrm{patch}_i(x)\in\mathbb{R}^{P^2C}$1) and higher spatial resolution, increasing computational cost. Depth (number of layers) is the most effective axis for scaling model capacity, followed by increasing $\mathrm{patch}_i(x)\in\mathbb{R}^{P^2C}$2. Reducing $\mathrm{patch}_i(x)\in\mathbb{R}^{P^2C}$3 also yields accuracy gains without more parameters.

4. Training and Transfer Protocols

ViT models are pre-trained on large datasets: ImageNet-21k (14M images, 21k classes) and JFT-300M (303M images, 18k classes, de-duplicated vs. downstream sets). The optimization strategy employs Adam ($\mathrm{patch}_i(x)\in\mathbb{R}^{P^2C}$4), weight decay of ≈0.1, and large batch size (4096). Training uses a warm-up period, followed by linear or cosine learning rate decay, and dropout ≈0.1 on MLPs (optionally none for the largest models).

Fine-tuning is performed by discarding the pre-trained classification head and adding a zero-initialized linear layer ($\mathrm{patch}_i(x)\in\mathbb{R}^{P^2C}$5, for $\mathrm{patch}_i(x)\in\mathbb{R}^{P^2C}$6 classes). Typically, fine-tuning operates at higher input resolution (384–512 px), 2D-interpolating the positional embeddings. Fine-tuning uses SGD with momentum 0.9, typical batch sizes of 512, cosine learning rate decay, and no weight decay. Standard hyperparameter grids are applied for learning rate search.

ImageNet fine-tuning occurs over 20,000 steps; smaller target sets use 500–10,000 steps. When pre-trained only on ImageNet (1.3M images), ViT-Large models underperform relative to CNNs. Pre-training on larger datasets (IN-21k, JFT-300M) is critical for strong performance (Dosovitskiy et al., 2020).

5. Performance and Computational Characteristics

Empirical evaluations demonstrate:

• ViT-H/14 achieves 88.55% top-1 on ImageNet, 90.72% on Imagenet-Real, 94.55% on CIFAR-100, and 77.63% on VTAB-19 after JFT-300M pre-training.
• ViT-L/16 achieves 87.76% on ImageNet, surpassing ResNet152×4 (BiT-L, 87.54%) while utilizing only 0.68k TPUv3-core-days (vs. 9.9k for BiT-L).
• ViT-L/16 pre-trained only on ImageNet-21k achieves 85.30% top-1 in 0.23k core-days.

ViT models require 2–4× fewer pre-training FLOPs than comparably accurate ResNets (R50–R200×3). Hybrid ViT+CNN architectures offer advantages only at smaller compute budgets. Inference speed on TPUv3 is comparable to ResNets, with higher memory efficiency, permitting larger per-core batch sizes (Dosovitskiy et al., 2020).

6. Key Insights: Data, Inductive Bias, and Attention

The data scale used for pre-training is critical. On large datasets (14M–300M images), ViT models rapidly surpass CNNs, indicating that scale outweighs explicit convolutional inductive bias. Overfitting arises on small datasets, confirmed by few-shot linear probe experiments.

Model performance correlates more with total pre-training compute than with parameter count. Depth scaling (up to 32 layers, with diminishing returns beyond 64) is especially effective. Visualizations of self-attention show that some heads operate globally from early layers, while others are local; the receptive field size grows systematically with depth. Attention rollout analyses reveal that the $\mathrm{patch}_i(x)\in\mathbb{R}^{P^2C}$7-to-patch attention vector localizes on semantically salient regions.

Positional embeddings are critical: simple learnable 1D embeddings suffice, with no observed benefit for 2D or relative positional embeddings. All positional encoding variants outperform models without positional bias by a wide margin.

Hybrid models, combining a shallow ResNet stem with ViT, are beneficial primarily at low compute but offer no significant advantage when scaling to large data and model sizes.

Preliminary masked-patch prediction (BERT-style) self-supervision achieves 79.9% top-1 for ViT-B/16 on ImageNet (vs. 75.9% from scratch), but remains ∼4% below supervised pre-training (Dosovitskiy et al., 2020).

7. Significance and Future Directions

ViT demonstrates that a pure Transformer architecture, with minimal vision-specific inductive bias, can achieve and surpass state-of-the-art accuracy in image classification, provided access to large-scale training data and compute. This finding challenges the necessity of convolutional priors in computer vision, and enables Transformer-only architectures for tasks beyond classification, such as detection, segmentation, and self-supervision. A plausible implication is that Transformer-based architectures, with sufficient scale, may continue to subsume traditionally convolutional approaches as resources and datasets grow (Dosovitskiy et al., 2020).