Rupesh Kumar Srivastava
Klaus Greff
̈
J urgen
Schmidhuber
The Swiss AI Lab IDSIA / USI / SUPSI
{rupesh, klaus, juergen}@idsia.ch
Abstract
Theoretical and empirical evidence indicates that the depth of neural networks
is crucial for their success. However, training becomes more difficult as depth
increases, and training of very deep networks remains an open problem. Here we
introduce a new architecture designed to overcome this. Our so-called highway
networks allow unimpeded information flow across many layers on information
highways. They are inspired by Long Short-Term Memory recurrent networks and
use adaptive gating units to regulate the information flow. Even with hundreds of
layers, highway networks can be trained directly through simple gradient descent.
This enables the study of extremely deep and efficient architectures.
1
Introduction & Previous Work
Many recent empirical breakthroughs in supervised machine learning have been achieved through
large and deep neural networks. Network depth (the number of successive computational layers) has
played perhaps the most important role in these successes. For instance, within just a few years, the
top-5 image classification accuracy on the 1000-class ImageNet dataset has increased from ∼84%
[1] to ∼95% [2, 3] using deeper networks with rather small receptive fields [4, 5]. Other results on
practical machine learning problems have also underscored the superiority of deeper networks [6]
in terms of accuracy and/or performance.
In fact, deep networks can represent certain function classes far more efficiently than shallow ones.
This is perhaps most obvious for recurrent nets, the deepest of them all. For example, the n bit
parity problem can in principle be learned by a large feedforward net with n binary input units, 1
output unit, and a single but large hidden layer. But the natural solution for arbitrary n is a recurrent
net with only 3 units and 5 weights, reading the input bit string one bit at a time, making a single
recurrent hidden unit flip its state whenever a new 1 is observed [7]. Related observations hold for
Boolean circuits [8, 9] and modern neural networks [10, 11, 12].