Semi-Supervised & Self-SupervisedSelf-Supervised LearningGenerative Modeling

Masked Autoencoders (MAE)

Primary task · Generative Modeling

Masked Autoencoders (MAE) is a semi-supervised & self-supervised method in the self-supervised learning family. This page summarizes its mechanism, practical uses, important trade-offs, and a browser-based concept explorer.

← Directory

Visual intuition

From data to learned behaviour

Masked Autoencoders (MAE) converts patterns in observed data into a reusable prediction or representation rule. The most useful way to understand it is to watch what internal structure changes during training and how that learned structure changes outputs.

Infographic

Conceptual simulation

Watch the learning mechanism form

The structure below is synchronized with the same training state used by the prediction simulation.

Mechanism view

Training control centre

Control both simulations together

Reset regenerates the synthetic data and model state. Train animates to completion. Pause freezes the animation. Train Step advances one learning stage.

Step 0 / 12

Model simulation

Inspect the learned prediction / representation

Synthetic data are generated locally in your browser.

Model description

Understand Masked Autoencoders (MAE) after watching it learn

This section connects the animation to the actual statistical or computational idea behind the model.

Deep description

Masked Autoencoders (MAE) Masked Autoencoders (MAE) is a semi-supervised & self-supervised method in the self-supervised learning family. This page summarizes its mechanism, practical uses, important trade-offs, and a browser-based concept explorer.

What is learned. During training, the algorithm builds or adjusts the parameters and internal representation used by Masked Autoencoders (MAE). The core learning mechanism is: Masks a high percentage (75-80%) of input image patches and trains an asymmetric Vision Transformer encoder-decoder to reconstruct the missing pixels.

How training becomes inference. Prepare data → initialise the model state → evaluate the current objective → update parameters or structure → validate progress → use the final state for inference. Once training stops, the fitted state is reused on unseen inputs rather than being reconstructed from scratch. The resulting output is: New samples or reconstructed/denoised representations drawn from the learned data distribution.

Why practitioners use it. Massive training speedup (encoder processes only 20-25% unmasked tokens); scales effectively to billion-parameter models. Typical fits include Foundation vision model pre-training, satellite image analysis, medical scan representations.

What to verify before trusting it. Primarily visual; requires substantial compute budgets and vision transformer architecture. The visual simulation is intentionally simplified, so real use should still validate preprocessing, data independence, hyperparameters, uncertainty and task-appropriate metrics.

Internal statethe parameters and internal representation used by Masked Autoencoders (MAE)

Typical outputNew samples or reconstructed/denoised representations drawn from the learned data distribution.

Good fitFoundation vision model pre-training, satellite image analysis, medical scan representations.

Main cautionPrimarily visual; requires substantial compute budgets and vision transformer architecture.

1Training data→

2Learning objective→

3Internal model state→

4Prediction / representation→

5Evaluation

Intuition

What the model is trying to learn

Mathematical lens

Core logic

Masks a high percentage (75-80%) of input image patches and trains an asymmetric Vision Transformer encoder-decoder to reconstruct the missing pixels. The mathematical objective determines which model states are considered better, while regularisation and validation constrain how much complexity should be trusted.

Training sequence

How learning progresses

Prepare data → initialise the model state → evaluate the current objective → update parameters or structure → validate progress → use the final state for inference.

Original mechanism

Taxonomy description

Masks a high percentage (75-80%) of input image patches and trains an asymmetric Vision Transformer encoder-decoder to reconstruct the missing pixels.

Evaluation guide

How to evaluate this model responsibly

ValidationChoose validation that matches the independence assumptions of the data.

MetricsUse task-specific primary and complementary metrics.

HPOEstablish a baseline first, then search the parameters that materially change capacity.

Post-processingValidate any downstream transformation on held-out data.

Hyperparameters

Key parameters

mask_ratioTypical: 0.75

Fraction of patches hidden from the encoder.

patch_sizeTypical: 16

Spatial patch size.

encoder_depthTypical: 12+

Transformer encoder layers.

decoder_dimTypical: 512

Lightweight decoder width.

Use & trade-offs

Where it fits

Typical applications

Foundation vision model pre-training, satellite image analysis, medical scan representations.

Strengths

Massive training speedup (encoder processes only 20-25% unmasked tokens); scales effectively to billion-parameter models.

Limitations

Primarily visual; requires substantial compute budgets and vision transformer architecture.

Code example

Minimal Python implementation

import torch
import torch.nn as nn

torch.manual_seed(7)
x=torch.randn(16,12); mask=torch.rand_like(x)<0.5; visible=x.masked_fill(mask,0)
mae=nn.Sequential(nn.Linear(12,20),nn.ReLU(),nn.Linear(20,12)); recon=mae(visible)
loss=((recon-x)[mask]**2).mean()
print("STEP 1 · Mask part of each input and keep the visible values")
print("STEP 2 · Reconstruct the complete input")
print("STEP 3 · masked fraction", round(mask.float().mean().item(),3), "masked MSE", round(loss.item(),4))

Expected / representative output

STEP 1 · Mask part of each input and keep the visible values
STEP 2 · Reconstruct the complete input
STEP 3 · masked fraction 0.542 masked MSE 0.9357