A (good) representation is easy to exploit.

Facial Action Classification

• hard to get sufficient amount of labeled data
• finding representations unsupervised
• profit ?

First Impression

from Kamil Czarnogórski

int8.io/variational-autoencoder-in-tensorflow

Autoencoder

trained on one face

“Picture of a face”

trained on multiple faces

“Picture of a X face”

↓

Autoencoder

↓

Representation

=

“XnXeaXrXlXy” FXA CXoXdXe

The solution

Probabilistic Models

• model (e.g. ANN) that acts like a certain p.d.f.
• naive model wont work

Generative Adversarial Network

Variational Autoencoder

Adversarial Autoencoder

AAE training

Grouped observations

Batch evaluation

Batch evaluation disentangled

Hijacking the style

Setup

• Quantitative: classifying CMNIST (4 setups)
• Qualitative: manipulating representations, then reconstructing

Q1: full classifier

Q2: classifier on full entangled representation

Q3/Q4: classifier on each representation

Results

Training error

Recognition rate

Reconstruction

Latent cross product

Random style

Random content

Conclusions

• It lives !
• stripping the style improves classification performance
• style holds (nearly) no information on content
• no dependency on each other when reconstruction with recombination
• non proportional error, when randomizing style

Outlook, discussion

• More complex dataset: e.g. YTFDB
• better classifier: may outperform Q1 ?
• reconstruction imbalance: learning defect or feature complexity difference ?
• any difference with an explicit variance regularizer on style ?
• comparing to VAE that works on parameters
• β-VAE + this.algorithm: get they along ?

A failed try on YTFDB

Notable Literature

D. Bouchacourt, R. Tomioka, and S. Nowozin: ‘Multi-level variational autoencoder: Learning disentangled representations from grouped observations’, 2017

Thank you for your attention !