
Exploring Randomly Wired Neural Networks for Image Recognition
Saining Xie Alexander Kirillov Ross Girshick Kaiming He
Facebook AI Research (FAIR)
Abstract
Neural networks for image recognition have evolved
through extensive manual design from simple chain-like
models to structures with multiple wiring paths. The suc-
cess of ResNets [11] and DenseNets [16] is due in large
part to their innovative wiring plans. Now, neural architec-
ture search (NAS) studies are exploring the joint optimiza-
tion of wiring and operation types, however, the space of
possible wirings is constrained and still driven by manual
design despite being searched. In this paper, we explore a
more diverse set of connectivity patterns through the lens of
randomly wired neural networks. To do this, we first define
the concept of a stochastic network generator that encap-
sulates the entire network generation process. Encapsula-
tion provides a unified view of NAS and randomly wired net-
works. Then, we use three classical random graph models
to generate randomly wired graphs for networks. The re-
sults are surprising: several variants of these random gen-
erators yield network instances that have competitive ac-
curacy on the ImageNet benchmark. These results suggest
that new efforts focusing on designing better network gen-
erators may lead to new breakthroughs by exploring less
constrained search spaces with more room for novel design.
1. Introduction
What we call deep learning today descends from the
connectionist approach to cognitive science [38, 7]—a
paradigm reflecting the hypothesis that how computational
networks are wired is crucial for building intelligent ma-
chines. Echoing this perspective, recent advances in com-
puter vision have been driven by moving from models with
chain-like wiring [19, 53, 42, 43] to more elaborate connec-
tivity patterns, e.g., ResNet [11] and DenseNet [16], that are
effective in large part because of how they are wired.
Advancing this trend, neural architecture search (NAS)
[55, 56] has emerged as a promising direction for jointly
searching wiring patterns and which operations to per-
form. NAS methods focus on search [55, 56, 33, 26, 29,
27] while implicitly relying on an important—yet largely
overlooked—component that we call a network generator
(defined in §3.1). The NAS network generator defines a
family of possible wiring patterns from which networks
classifier classifier classifier
conv
1
conv
1
conv
1
Figure 1. Randomly wired neural networks generated by the
classical Watts-Strogatz (WS) [50] model: these three instances
of random networks achieve (left-to-right) 79.1%, 79.1%, 79.0%
classification accuracy on ImageNet under a similar computational
budget to ResNet-50, which has 77.1% accuracy.
are sampled subject to a learnable probability distribution.
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. Given this perspective, we ask: What
happens if we loosen this constraint and design novel net-
work generators?
We explore this question through the lens of randomly
wired neural networks that are sampled from stochastic
network generators, in which a human-designed random
process defines generation. To reduce bias from us—the
authors of this paper—on the generators, we use three clas-
sical families of random graph models in graph theory [51]:
the Erd
˝
os-R
´
enyi (ER) [6], Barab
´
asi-Albert (BA) [1], and
Watts-Strogatz (WS) [50] models. To define complete net-
works, we convert a random graph into a directed acyclic
graph (DAG) and apply a simple mapping from nodes to
their functional roles (e.g., to the same type of convolution).
1
arXiv:1904.01569v2 [cs.CV] 8 Apr 2019