Towards Better Analysis of

Deep Convolutional Neural

Networks
 Neural Network
Neural network" (NN), is a learning
algorithm that is inspired by the
structure and functional aspects of
biological neural networks.

https://www.youtube.com/watch?v=gc
K_5x2KsLA
 ANN & CNN
The main difference between ANN and
CNN is that the convolutional neural
network (CNN) has layers
of convolution and pooling.

This means that the first layers that come

after the input, do not use all input features
at the same time but rather features that
are "connected".
 Gist
Deep convolutional neural networks
(CNNs) have achieved breakthrough
performance in many pattern recognition
tasks such as image classification.

But, the development of high-quality deep

models typically relies on a substantial
amount of trial-and-error
 Gist
As there is still no clear understanding of
when and why a deep model works.

A visual analytics approach is needed for

better understanding, diagnosing, and
refining deep CNNs

This approach formulates a deep CNN as a

directed acyclic graph
 Gist
Based on this graph, a hybrid visualization
is developed to disclose the multiple facets
of each neuron and the interactions
between them.

A hierarchical rectangle packing algorithm

and a matrix reordering algorithm to show
the derived features of a neuron cluster.
 Gist
A biclustering-based edge bundling
method to reduce visual clutter caused by
a large number of connections between
neurons.

The method is evaluated on a set of CNNs

and the results are favourable.
 Introduction
Deep convolutional neural networks (CNNs) have
demonstrated significant improvements over
traditional approaches in many pattern
recognition tasks.

Recently, deep CNNs have been employed to

extract robust representations and help make
decisions.

This has led to human-level performance in

intelligent tasks such as Atari games and the
game of Go.
 Introduction
But, a deep CNN is often treated as a black box
model because of its incomprehensible functions
and unclear working mechanism.

There are two technical challenges.

A CNN may consist of tens or hundreds of layers,
with thousands of neurons (width) in each layer,
as well as millions of connections between
neurons.

A CNNs consist of many functional components

whose values and roles are not well understood
either as individuals or as a whole.
 Introduction
In addition, the interaction between non-linear
components with each other and with other linear
components in a CNN is not well understood by
experts.

In most cases, it is hard to summarize reusable

knowledge from a failed or successful training
case and transfer it to the development of other
relevant deep learning models
 CNNVis
A visual analytics system called CNNVis is
developed.

This aims to help machine learning experts better

understand, diagnose, and refine CNNs.

In CNNVis, diagnosis helps experts analyze a

training process that has failed to converge.

The refinement aims to find a potential direction

to improve model accuracy.
CNNVis
 CNNVis
A directed acyclic graph (DAG), in which each
node represents a neuron and each edge
represents the connection between neurons.

In order to visualize a large CNN, the layers in the

network are clustered and representative one
from each layer cluster is selected.

The neurons in each representative layer is

clustered and several representative neurons
from each neuron cluster is selected
 CNNVis
Key Contributions:

A visual analytics system that helps experts

understand, diagnose, and refine deep CNNs.

A hybrid visualization that combines a DAG with

rectangle packing, matrix visualization, and a
biclustering-based edge bundling method.
 CNN Background
The CONVOLUTION layers parameters consist
of a set of learnable filters.

Every filter is small spatially, but extends through

the full depth of the input volume.

Pooling function is to progressively reduce the

spatial size of the representation to reduce the
amount of parameters and computation in the
network.
 CNN Background
The convolutional layer aims to detect local
combinations of features from the previous layer

The pooling layer aims to merge semantically

similar features into one.
 Visualization - Hierarchical
Rectangle packing
Step 1: Hierarchical clustering.

A hierarchical clustering is done to divide the

image patches into several groups.
It is start with the cluster containing all of the
neurons.
Then we repeatedly split a cluster until the
number of neurons in it is smaller than a
threshold.
Visualization - Hierarchical
Rectangle packing
 Visualization - Hierarchical
Rectangle packing
Step 2: Computing the layout area for each cluster.
Based on the hierarchical clustering results, we
compute the layout area for each sub-cluster
using a Tree-map layout algorithm.

Step 3: Rectangle packing of each cluster.

The position and size for each image patch is
computed using the state-of-the-art rectangle
packing algorithm.
Visualization - Matrix Reordering

The order of columns should be consistent in

different neuron clusters.

Otherwise, experts are unable to directly compare

the roles of neurons in two neuron clusters
because of the different order of columns.

As a solution, reordering of the rows (neurons) in

the matrix is performed.
Visualization - Matrix Reordering
The basic idea of this algorithm is to maximize the
sum of the similarities between adjacent neurons
in the matrix.

The aim is to place neurons with similar

activations close to each other, thus revealing the
cluster pattern in the neuron cluster
 Visualization - Biclustering-based
Edge Bundling
Initially each edge was visualized as a curve
which resulted in visual clutter caused by millions
of edges between nodes.

In order to reduce the clutter, the geometry-based

edge bundling method was used to cluster the
edges between two layers.

This bundling method reduces visual clutter to

some extent.
 Visualization - Biclustering-based
Edge Bundling
But, the clusters revealed by the geometry-based
bundling methods did not help their analysis
because the edges with similar weights were not
clustered together.

The edges with larger absolute weights are

important, because this indicates that the
corresponding inputs have a larger impact on the
output.
 Visualization - Biclustering-based
Edge Bundling
To fulfil this requirement, a biclustering-based
edge bundling method is deployed to bundle
edges with both similar and large absolute
weights.

For a given layer, a bicluster is a subset of input

neuron clusters and a subset of output neuron
clusters.
 Visualization - Biclustering-based
Edge Bundling
This method can logically aggregate multiple
individual connections and thus provides an
opportunity to visually bundle edges between
neuron clusters.

In a CNN, each edge has a weight and the edges

with similar weights has to be bundled.
Visualization - Biclustering-based
Edge Bundling
Results
Results
 Conclusion
The IEEE paper presented a novel visual
analytics system to help machine learning experts
better understand, diagnose, and refine CNNs.

This system is powered by a hybrid visualization

consisting of rectangle packing, matrix ordering,
and biclustering-based edge bundling.
 Conclusion
It allows experts to explore and understand a
deep CNN from different perspectives.

In addition, it enables experts to diagnose and

refine the CNN architecture to further improve
performance.
 Reference
http://cs231n.github.io/convolutional-networks/
https://www.youtube.com/watch?v=gcK_5x2KsLA
Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner.
Gradient-based learning applied to document
recognition. IEEE, 86(11):22782324, 1998.
C. Li, J. Zhu, T. Shi, and B. Zhang. Max-margin
deep generative models. In NIPS, pages 1828
1836, 2015.