Representation Learning with Contrastive Predictive Coding

[flrngel]

https://arxiv.org/abs/1807.03748 ~big fan of Aaron van den Oord~

Abstract

• propose universal unsupervised learning approach to extract useful representations from high-dimensional data, CPC(Contrastive Predictive Coding)
• use probabilistic contrastive loss which induce latent space to capture information that is useful to predict future samples and tractable by using negative sampling

1. Introduction

• unsupervised learning is an important stepping stone towards robust and generic representation learning
• idea of predictive coding theories comes from neuroscience, Word2Vec, colorization, etc.
• paper proposes
  • compress high-dimensional data into compact latent embedding space which conditional predictions are easier to model
    1. predict further steps with autoregressive model
    2. use NCE loss
    3. train CPC as end-to-end resulting model
• CPC model outperforms other models in various of domains

2. Contrastive Predicting Coding

2.1. Motivation and Intuitions

• in time series and high dimension modeling, they use next step prediction exploit the local smoothness of signal
  • further in future, shared information becomes much lower and model needs to infer more global structure and these are called 'slow features'
• 'class size' < 'latent variable' in information

2.2. Contrastive Predictive Coding

• g_enc maps the input sequence of observations x_t to latent representations z_t = g_enc(x_t)
• autoregressive model gar summarizes all `z<=tin lthe latent space and produces a context latent representationc_t = gar(z<=t)`

it means

• proposed model, either z_t and c_t could be used as representation for downstream tasks
• c_t can be used if extra content from the past is useful
• z_t might not contain enough information to capture phonetic content

2.3. Noise Contrastive Estimation Loss

2.4. Related work

• triplet losses using max-margin to separate positive from negative examples
• time contrastive learning which minimize between embeddings from multiple viewpoints of the same scene and maximize different embeddings extracted from different timesteps
• in word2vec neighbouring words are predicted using a contrastive loss

3. Experiments

3.1. Audio

3.2. Vision

model

experiment result

3.3. Reinforcement Learning

4. Conclusions

• CPC combines autoregressive modeling and noise-contrastive estimation with intuitions from predictive coding to learn abstract representations in an unsupervised fashion

Notes

• best paper of this week
• I will implement this paper to code as soon as possible

[paganpasta]

Hey @flrngel ! Thanks for the notes and the corresponding repository. I was trying to understand the statement made by the authors which you cite in Section 2.1(last paragraph). The statement reads " ... by maximising the MI between the encoded signals( which is bounded by the MI between the input signals)... Could you please elaborate on it? Does it mean I(g(x);g(x+t)) < I(x; x+t), if so why ? Sorry for a naive question.

[flrngel]

@MacroMayhem Equation (2) and Appendix A will help you to undertand I guess.