# Unsupervised learning can learn from features without label

Last updated on：9 months ago

Unsupervised learning can learn features by specifically designed loss **without any pre-labels**. There are two most common seen unsupervised learning methods, which are **autoencoder and GANs**. In this article, we mainly focus on autoencoder.

# Introduction

Unsupervised learning is about problems where we **don’t have labelled answers,** such as clustering, dimensionality reduction, and anomaly detection.

Unsupervised learning is usually employed (e.g., clustering) to **separate the images into two sets or clusters based on some inherent features** of the pictures like colour, size, shape, etc.

## Motivation

- It is
**not easy to obtain a large amount of labelled data**in practice, but it is easy to get a large amount of unlabelled data. **Learning a suitable feature extractor using unlabelled data**and then learning the classifier using labelled data can improve the performance.

## From the data point of view

Data in both input $x$ and output $y$ (learn the mapping $f$)

# Autoencoder

Autoencoder would be an unsupervised learning method if **we considered the features the “output”**. Autoencoder is also a self-taught learning method, a type of artificial neural network used to learn efficient coding of **unlabelled data.** The encoding is validated and refined by attempting to **regenerate the input from the encoding**.

**Word2Vec** is another unsupervised, self-taught learning example.

Given M data samples,

$$\mathcal{L}_{\text{MSE}} = \frac{1}{M} \sum _{m=1}^M ||\hat{x}^m - x^m|| ^2_2$$

It is trying to **learn an approximation to the identity function** so that the input is “**compress**“ to the **“compressed” features**, discovering interesting structure about the data.

## Sparse autoencoder

Even when the number of hidden units is large (perhaps even more significant than the number of input pixels), we can still discover an interesting structure, by **imposing other constraints on the network**.

In particular, if we impose a “**sparsity” constraint** on the hidden units, then the autoencoder will still discover exciting **structures in the data**, even if the number of hidden units is significant.

One neuron = one feature extractor.

Given M data samples and Sigmoid activation function, the active ratio of a neuron $a_j$:

$$\hat{\rho} _j = \frac{1}{M}\sum _{m=1}^{M}a_j$$

This makes the output “sparse”. We would like to **enforce the following constraint**, where $\rho$ is a “**sparsity parameter**“, such as 0.2 (20% of the neurons)

$$\hat{\rho}_j = \rho$$

The penalty term is as follow, where $s$ is the number of output neurons.

$$\mathcal{L}=\sum^S _{j=1}KL(\rho||\hat{\rho} _j)=\sum^S _{j=1} (\rho\text{log} \frac{\rho}{\hat{\rho} _j} + (1-\rho)\text{log}\frac{1-\rho}{1-\hat{\rho} _j}$$

The total loss:

$$\mathcal{L} _{\text{total}} = \mathcal{L} _{\text{MSE}} + \mathcal{L} _{\rho}$$

Method | Hidden activation | Reconstruction activation | Loss function |
---|---|---|---|

method1 | Sigmoid | Sigmoid | $$\mathcal{L} _{\text{total}} = \mathcal{L} _{\text{MSE}} + \mathcal{L} _{\rho}$$ |

method2 | ReLU | Softplus | $$\mathcal{L}_{\text{total}} = \mathcal{L} _{\text{MSE}} + ||\bf{a}||$$ |

$||\bf{a}||$ is $\mathcal{L}_1$ on the hidden activation output.

## Other types

**Denoising autoencoder**

**Stacked autoencoder**

**Variational autoencoder**

# Applications

## Learn segmentation via domain adaptation

Each encoder receives a different augmented version of the input image, generated by $f_T$. The loss:$$\mathcal{L}_{\text{SRMA}} = \mathcal{L} _{\text{IND}} + \mathcal{L} _{\text{CRD}}$$

$ \mathcal{L} _{\text{IND}} $ is the in-domain loss, and $\mathcal{L} _{\text{CRD}}$ is the cross-domain loss.

In Abbet et al. work, the domain adaptation is an unsupervised process.

**Two other predictions** generate target pseudo labels for each segmentation network generated by during pseudo labelling.

Implicit pseudo supervision includes semantic feature alignment (SFA) and adaptation ability estimation (AAE). AAE exploits the **adaptation ability for each pixel** and each network and rectifies the pseudo labels to get $\mathcal{L}_{\text{aae}}$.

Semantic feature alignment (SFA) **minimizes the distance of feature centroids** between the same class for background categories to get $\mathcal{L}_{\text{back}}$ and maximizes the distance between different categories for foreground types to get $\mathcal{L}_{\text{fore}}$.

## Learn segmentation via clustering

SegDiscover has three steps:

**Creating concept primitives**. The images in the dataset are segmented into “concept primitives” (patches/super pixels)**Concept discovery**via primitive**clustering**. The primitive patches are fed into a pretrained self-supervised network, and potential concepts are discovered by clustering in its latent space.**Concept refinement**via neural**network smoothing**, the cluster labels are**mapped back to the original images**to form pseudo segmentation labels. A segmentation network is trained to predict the pseudo labels given the authentic images. The segmentation network refines the learned concepts.

The pooled tensor is **reshaped and fed into the feature encoder** to obtain latent representations. Features (blue points) are clustered in the latent space to get cluster centres through GMM (feature prototypes, shown as red points), and clustering loss can be calculated.

## Cluster assignment

In contrastive learning methods applied to instance classification, the features from different transformations of the same images are compared directly. SwAV (**s**wapping **a**ssignments between multiple **v**iews of the same image) obtain “code” by **assigning features to prototype vectors**. And then, a swapped prediction trick is applied where the codes received from **one data augmented view are predicted using the other view**.

# Reference

[1] Hao Dong, Learning Methods.

[2] Abbet, C., Studer, L., Fischer, A., Dawson, H., Zlobec, I., Bozorgtabar, B. and Thiran, J.P., 2022. Self-Rule to Multi-Adapt: Generalized Multi-source Feature Learning Using Unsupervised Domain Adaptation for Colorectal Cancer Tissue Detection. Medical Image Analysis, p.102473.

[3] Xu, W., Wang, Z. and Bian, W., 2022. Unsupervised Domain Adaptation with Implicit Pseudo Supervision for Semantic Segmentation. arXiv preprint arXiv:2204.06747.

[4] Huang, H., Chen, Z. and Rudin, C., 2022. SegDiscover: Visual Concept Discovery via Unsupervised Semantic Segmentation. arXiv preprint arXiv:2204.10926.

[5] Liu, J., Qi, X., Su, S., Prescott, T. and Sun, L., 2021, March. Zero-shot Anomalous Object Detection using Unsupervised Metric Learning. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2021) Proceedings. Sheffield.

[6] Dike, H.U., Zhou, Y., Deveerasetty, K.K. and Wu, Q., 2018, October. Unsupervised learning based on artificial neural network: A review. In 2018 IEEE International Conference on Cyborg and Bionic Systems (CBS) (pp. 322-327). IEEE.

[7] Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P. and Joulin, A., 2020. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33, pp.9912-9924.

本博客所有文章除特别声明外，均采用 CC BY-SA 4.0 协议 ，转载请注明出处！