Simple neural network

Markus Dreyer, Feb 2015

A feedforward neural network for tagging (i.e., sequence labeling).

This code demonstrates that implementing a neural network does not require much code. The implementation is simpler than CRFs or related models, since no manual feature engineering is required.

The implementation follows this project description.

Inspired by this post and others, I am using automatic differentiation here, rather than implement the gradients by hand. This allows for rapid prototyping and quick experimentation with different models. Automatic differentiation is different from and much faster than numerical differentiation. See here for more information.

The code depends on two external libraries:

• Eigen for matrix data structures and matrix multiplication.

• Adept for automatic differentiation.

Adept 1.0 (and earlier versions, presumably) have a bug in the tanh differentiation. To fix it, change line 1486 in adept.h from this:

ADEPT_DEFINE_UNARY_FUNCTION3(Tanh, tanh, Real e=exp(2.0*a_.value()), 4.0*e/((2.0+2)*e+1.0))

to this:

ADEPT_DEFINE_UNARY_FUNCTION3(Tanh, tanh, Real tmp=tanh(a_.value()), 1.0-tmp*tmp)

I hear that the bugfix will be included in the next Adept version.

To compile simple-nn, specify the Eigen and Adept locations in the Makefile, and type make. Then type:

./simple-nn -h

Which displays something like this:

./simple-nn: A neural network for tagging (experimental).
Options:
 -h      Show this help text.
 -d dir  Specify the data directory.
 -f [tanh|sigmoid|softplus] Specify the nonlinear function.
 -n num  Specify the window size.
 -H num  Specify the hidden layer size.
 -l num  Specify the gradient descent learning rate.
 -s      Skip test and evaluation at the end.

Simply run with defaults like this:

./simple-nn

Caveats

It seems that Adept slows down the score computation during training considerably, so this is currently good for prototyping purposes only.