PR-155: Exploring Randomly Wired Neural Networks for Image Recognition

Exploring Randomly Wired Neural Networks
for Image Recognition
SainingXie, et al., “Exploring RandomlyWired Neural Networks for Image Recognition”, arxiv:1904.01569
Facebook AI Research(FAIR)
7th April, 2019
PR12 Paper Review
JinWon Lee
Samsung Electronics

Introduction
• What we call deep learning today descends from the connectionist approach to
cognitive science.
• How computational networks are wired is crucial for building intelligent
machines.
• ResNet and DenseNet, that are effective in large part because of how they are
wired.
• Advancing this trend, neural architecture search (NAS) has emerged as a
promising direction for jointly searching wiring patterns and which operations to
perform.

Introduction
• However, like the wiring patterns in ResNet and DenseNet, the NAS network
generator is hand designed and the space of allowed wiring patterns is
constrained in a small subset of all possible graphs.
• What happens if we loosen this constraint and design novel network generators?
<NASNet>

MainTopic
• Design a Network Generator not an Individual Network!
• The authors suggest a new transition from designing an individual network to
designing a network generator may be possible, analogous to how our
community have transitioned from designing features to designing a network
that learns features.
• Rather than focusing primarily on search with a fixed generator, authors suggest
designing new network generators that produce new families of models for
searching.

RelatedWork
• Network Wiring
 CNN & RNN use chain-like wiring patterns.
 LSTM, Inception, ResNet, and DenseNet wiring patterns are effective in general.
• Neural Architecture Search(NAS)
 Recent research on NAS mainly focuses on optimization methods, including RL, progressive,
gradient-based, etc.
 The search space in these NAS works is largely unchanged in these works.

RelatedWork
• Randomly Wired Machines
 Turing suggested a concept of unorganized machines, which is a form of the earliest
randomly connected neural networks.
 Minsky and Rosenblatt also suggested randomly connected machines
• Relation to Neuroscience
 Turing analogized the unorganized machines to an infant human’s cortex.
 Rosenblatt pointed out that “at birth, the construction of the most important networks is
largely random.”
• Random Graphs in GraphTheory
 Random graphs are widely studied in graph theory.
 The definition of the random graph model determines the prior knowledge encoded in the
resulting graphs and may connect them to naturally occurring phenomena.

Network Generator
• Defining a network generator as a mapping 𝑔 from a parameter space Θ to a
space of neural network architecture Ν.
 The set Ν is typically a family of related networks, for example,VGG nets, ResNets, or
DenseNets.
• The parameters 𝜃 ∈ Θ specify the instantiated network and my contain diverse
information.
 For example, in a ResNet generator, 𝜃 can specify the number of stages, number of residual
blocks for each stage, depth/width/filter sizes, activation types, etc.

Network Generator
• Stochastic Network Generator
 The network generator 𝑔 𝜃 performs a deterministic mapping.
 We can extend 𝑔 to accept an additional argument 𝑠 that is the seed of a pseudo-random
number generator that is used internally by 𝑔.
 We call generators of the form 𝑔 𝜃, 𝑠 stochastic network generators.
• NAS from the Network Generator Perspective
 The weight matrices of the LSTM are the parameters 𝜃 of the generator.
 Given the probability distribution conditioned on 𝜃 and the seed 𝑠, each step samples a
construction action(e.g., insert an operator, connect two nodes).
 NAS is the network generator that is hand-designed and encodes a prior from human
knowledge.
 The network space Ν has been restricted by hand-designed rules. (e.g., 5 nodes in a cell
always have input degree 2 and output degree 1)

RandomlyWired Neural Networks
• To investigate how important the generator design is necessary to study new
network generators that are substantially different from the NAS generator.
• So, authors define network generators that yield networks with random graphs,
subject to different human-specific priors.
 They use three classical random graph models – ER, BA,WS models
• Generating General Graphs
 Starting by generating a general graph.
 Without restricting how the graphs correspond to neural networks.
 Once a graph is obtained, it is mapped to a computable neural networks.
 The mapping is in itself arbitrary, and authors intentionally use a simple mapping.

• Edge Operation
 The edges are data flow – a directed edge sends data from one node
to another node.
• Node Operation
 Aggregation
 The input data to a node are combined via a weighted sum; the weights
are learnable and positive.
 Transformation
 ReLU-convolution-BN triplet
 The same type of convolution is used for all nodes. – 3x3 separable
convolution (3x3 depthwise conv  1x1 pointwise conv)
 Distribution
 The same copy of the transformed data is sent out by the output edges of
the node.

• Nice Properties of Node Operation
 Additive aggregation maintains the same number of output channels as input channels.
 Transformed data can be combined with the data from any other nodes.
 Fixing the channel count then keeps the FLOPs and parameter count unchanged for each
node, regardless of its input and output degrees.
 Aggregation and distribution are almost parameter-free, regardless of input and output
degrees.
 Also, the overall FLOPs and parameter count of a graph are roughly proportional to the
number of nodes and nearly independent of the number of edges
 This enables the comparison of different graphs without inflating/deflating model
complexity. Differences in task performance are therefore reflective of the properties of
the wiring pattern.

• Input and Output Nodes
 A general graph is not yet a valid neural network.
 It may have multiple input and output nodes.
 Creating a single extra node that is connected to all original
input nodes.
 Unique input node.
 Similarly, creating a single extra node that is connected to all
original output nodes.
 Unique output node & computing the average(unweighted) from
all original output nodes.
extra input node
extra output node

• Stages
 In image classification in particular, it is common to
divide a network into stages that progressively down-
sample feature maps.
 Simple strategy
 An entire network consists of multiple stages.
 One random graph represents one stage, and it is
connected to its preceding/succeeding stage by its
unique input/output node.
 For all nodes that are directly connected to the input
node, their transformations are modified to have a
stride 2.
 The channel count in a random graph is increased by
2x when going from one stage to the next stage.
<RandWire Architectures>

conv2
conv3
conv4
conv5
stride 2 convolution

Random Graph Models
• Erdös-Rényi (ER), 1959.
 Has N nodes
 An edge between two nodes is connected with probability P.
 The ER generation model has only a single parameter P, and is denoted as ER(P).
 Any graph with N nodes has non-zero probability of being generated by the ER model.

Random Graph Models
• Barabási-Albert (BA), 1999.
 Generates a random graph by sequentially adding new nodes.
 The initial state is M nodes without any edges (1 ≤ M < N).
 Sequentially adds a new node with M new edges.
 For a node to be added, it will be connected to an existing node v with probability
proportional to v’s degree.
 The new node repeatedly adds non-duplicate edges in this way until it has M edges.
 Any graph generated by BA(M) has exactly M(N-M) edges.This gives one example on how
an underlying prior can be introduced by the graph generator in spite of randomness.
Initial state is not
the same as this paper

Random Graph Models
• Watts-Strogatz (WS), 1998.
 “SmallWorld” model.
 Stanley Milgram experiment, Erdös number, Bacon number(http://oracleofbacon.org/).
 High clustering, small diameter.
 N nodes are regularly placed in a ring and each node is connected to its K=2 neighbors on
both sides – starts from regular graph.
 Then, in a clockwise loop, for every node v, the edge that connects v to its clockwise i-th
next node is rewired with probability P.
 “Rewiring” is defined as uniformly choosing a random node that is not v and that is not a
duplicate edge.

Converting Undirected Graphs into DAGs
• Assign indices to all nodes in a graph and set the direction of every edge as
pointing from the smaller-index node to the larger-index one.  no cycle.
• Indexing
 ER – indices are assigned in a random order.
 BA – the initial M nodes are assigned indices 1 to M, and all other nodes are indexed
following their order of adding to the graph.
 WS – indices are assigned sequentially in the clockwise order.

Visualization of the Random Graphs regular graph

Design and Optimization
• Randomly wired neural networks are generated by a stochastic network
generator 𝑔 𝜃, 𝑠 .
• The random graph parameters, P, M, (K; P) in ER, BA,WS respectively, are part
of the parameters 𝜃.
• The “optimization” of such a 1- or 2-parameter space is essentially done by trial-
and-error by human designers. – line/grid search
• The accuracy variation of our networks is small for different seeds 𝑠 so they
perform no random search and report mean accuracy of multiple random
network instances.

Experiments
• Architecture Details
 A small computation regime – MobileNet & ShuffleNet
 A regular computation regime – ResNet-50/101
 N nodes, C channels determine network complexity.
 N = 32, C = 79 for the small regime.
 N = 32, C = 109 or 154 for the regular regime.
• Random Seeds
 Randomly sample 5 network instances, train them from scratch.
 Report the classification accuracy with “mean±std” for all 5 network instances.
• Implementation Details
 Train for 100 epochs
 Half-period-cosine shaped learning rate decay and initial learning rate 0.1
 The weight decay is 5e-5
 Momentum 0.9
 Label smoothing regularization with a coefficient of 0.1

Analysis Experiments
• All random generators provide decent accuracy over all 5 random network
instances. – no fails to converge.
• ER, BA, andWS all have certain setting that yield mean accuracy > 73%, with in a
<1% gap from the best accuracy fromWS(4, 0.75)
• The variation among the random network instances is low. (std : 0.2~0.4%)
• Different random generators may have a gap between their mean accuracies. It
means that random generator design plays an important role in the accuracy.

• Graph Damage.
 Graph damage by randomly removing one node
or edge.
 For networks generated byWS, the mean
degradation of accuracy is larger when the
output degree of the removed node is
higher(hub nodes).
 The accuracy loss is generally decreasing along
the x-asis in the right figure(bottom).

• Node Operations
 Almost all networks still converge to non-trivial results.
 The Pearson correlation between any two series in the below figure is 0.91~0.98.
 This suggests that the network wiring plays a role somewhat orthogonal to the role of the
chosen operations.

Comparisons
• Small Computation Regime
 250 epochs for fair comparisons.
 The mean accuracy achieved by RandWire is a competitive result, especially considering that
they perform no random search, and use a single operation type for all nodes.

Comparisons
• Regular Computation Regime
 Use a regularization method inspired by edge removal analysis. Randomly remove one edge
whose target node has input degree > 1 with probability of 0.1.
 Mean accuracies are respectively 1.9% and 1.3% higher than the ResNet-50 and ResNet-101
and are 0.6% higher than the ResNeXt.

Comparisons
• Larger Computation
 Same trained networks as regular computation regime, but only increase the test image size
to 320x320 without retraining.
 Mean accuracy is 0.7%~1.3% lower than the most accurate NAS results, but use only ~2/3
FLOPs and ~3/4 parameters. Networks are trained for 100 epochs and not on the target
image size.
• Object Detection
 The features learned by randomly wired networks can also transfer.

Conclusion
• Exploring randomly wired neural networks by three classical random graph
models from graph theory.
• The result were surprising: the mean accuracy of these models is competitive
with hand-designed and optimized from NAS.
• The authors hope that future work exploring new generator designs may yield
new, powerful networks designs.

PR-155: Exploring Randomly Wired Neural Networks for Image Recognition

More Related Content

What's hot (20)

Similar to PR-155: Exploring Randomly Wired Neural Networks for Image Recognition (20)

More from Jinwon Lee (8)

Recently uploaded (20)

PR-155: Exploring Randomly Wired Neural Networks for Image Recognition