論文分解器：Hintons capsule 2017

02-07

EN視頻筆記：草稿（可忽略），最後正式理解會用中文視頻筆記整理

CapsNet implementation

Abstract 視頻筆記中文 EN

what is a capsule?
what can this activity vector tell us?

how capsules cross level interact?
how to understand routing-by-agreement?

Introduction paragraph 1 視頻筆記中文 EN

how human vision ignore irrelevant details?
knowledge from a sequence of fixation points vs a single fixation point
Assumptions proposed in this paper

Introduction paragraph 2 視頻筆記 EN 中文

parse trees
parse tree vs neural network
capsules vs layer vs parse tree nodes
iterative routing process and problem to solve
help needed from pre-paper1,2

Introduction paragraph 3 視頻筆記 EN 中文

active capsule vs properties of an entity in an image
instantiation parameters vs image entity properties
existence as special property of the entity in image
orientation of vector represent all other properties of the entity in the image
non-linearity activation function vs capsule vector representation power

Introduction paragraph 4 視頻筆記 EN 中文

destination of a capsule
two things help capsule to fulfill this dream
dynamic routing mechanism: routing and dynamic
How the portion of each parent capsule vary?
how to measure similarity?

how to visualize this similarity comparison process?
what is "routing by agreement" mechanism?
What more power does "routing by agreement" offer to models?

Introduction paragraph 5 視頻筆記 EN 中文

What special about CNN?
What can CNN do with this specialty?
How capsule model inherit and differ from CNN?
What is special about capsule and capsule model?

section 2 paragraph 1-4 視頻筆記 EN 中文

Aim of the paper
how a capsule detect an entity?
current input, output vector, "squashing" function

total input
coupling coefficients
what is the relationship between $c_{ij}, b_{ij}$

section 2 paragraph 5-7 視頻筆記 EN 中文

what are $b_{ij}$ ?
how to update $b_{ij}$ ?
how to understand the pseudo code of Routing algo?

section 3 視頻筆記 EN 中文

probability existence of a capsules entity
how to design a loss for capsule?
confident on both yes and no

section 4-4.1 視頻筆記 EN 中文

How to construct a CapsNet?

How to train CapsNet to reconstruct digit image?
why scale down reconstruction loss when calc total loss?
How to make CapsNet to reconstruct a digit image?
from Fig3 can we see how good CapsNet really is?

section 5 視頻筆記 EN 中文

dataset
base line model
CapsNet
How comparison table show the effectiveness of routing and reconstruction?
How reconstruction help performance?

section 5.1-2 視頻筆記 EN 中文

What the individual dimensions of a capsule represent

how mask work
what reconstruction is actually learning
perturbed vector + decoder network help see what individual dimension represent?
Robustness to Affine Transformation
Compare CNN vs CapsNet on affNIST dataset

section 6-6.2 視頻筆記中文

dynamic routing vs overlapping objects
How to create overlapping MNIST, multi-MNIST?
How well CapsNet is doing with multi-MNIST?

section 7 視頻筆記中文

CapsNet 嘗試了其他幾個數據（都比較小）
在CIFAR10上的效果，只能等同於早起的CNN的效果

效果不佳的原因簡要分析

section8 視頻筆記中英文

早起speech recognition多用HMM模型
為什麼RNN比HMM更高效？one-to-n vs distributed representation
RNN 比 HMM高效多少？2N的參數 VS N*2的參數