Bi-Directional Block Self-Attention for fast and memory-efficient sequence modeling

[flrngel]

Paper at ICLR2018 https://openreview.net/forum?id=H1cWzoxA-
aka BiBloSAN (Bi-BloSAN)

Abstract

• CNN focuses local dependency
• Bi-BloSAN achieves better efficiency-memory trade-off than RNN/CNN/SAN models

1. Introduction

• RNN has problem for parallel processing
• CNN number growth by position distance
  • Convolution Seq2Seq: Linearly
  • ByteNet: Logarithmically
• Bi-BloSAN succeeds,
  • Transformer
  • DiSAN (Directional Self-Attention Network)
    • forward/backward mask
    • using feature level
• SAN(Self-Attention Network) has memory problem and Bi-BloSAN solves it
• Core idea of Bi-BloSAN is processing tokens after splitting sequence to same length
• Bi-BloSAN benefits: memory-efficient, fast (intra -> inter block makes this possible)

2. Background

2.2 Vanilla Attention and Multi-dimensional Attention

• Additive attention is better than Multiplicative(dot-product) attention for
  • memory, time, computing performance

2.3 Two types of Self-attention

• token2token
  • x_i, x_j uses same x
• source2token
  • token to entire sentence

2.4 Masked Self-attention

• (Shen et al., 2017) used mask to make one way direction
• M is for masking, uses (forward)
  • 0; if i<j
  • -inf; else

3. Proposed Model

m-BloSA

Bi-BloSAN

4. Experiments

Terminology

• x: input sequence
• q: query
• i: position from sequence
• P: feature importance with evidence