The Lens Neural Net Simulator

[Douglas Rohde]

                  Lens 2.3 is Now Available!!
                http://www.cs.cmu.edu/~dr/Lens

I'm pleased to announce the first public release of the Lens
neural network simulator.  Lens is a general-purpose simulator
that runs on Windows as well as most Unix platforms.  It is
primarily designed for training feed-forward and recurrent
backpropagation networks, but can be easily adapted to handle
other architectures and currently includes deterministic
Boltzmann machines, Kohonen networks, and interactive-activation
models.

Lens provides a variety of graphical interfaces to aid in
executing common commands, visualizing unit and link values,
accessing internal parameters, and plotting data.  However, it
also includes a complete scripting language (Tcl), that enables
networks to be constructed quickly and easily.

Although Lens was designed with the serious modeler in mind, it
would actually be a good choice for use in teaching introductory
courses.  Basic networks can be created in just a few simple
commands and students will be up and running in minutes.

The following are some of the main advantages of Lens over other
neural network simulators I have used, including PDP++, SNNS,
Xerion, RCS, Tlearn, and Matlab:

-SPEED: Lens is up to 4 times faster than the other major
    simulators.  This is mainly due to tight inner loops that
    minimize memory references and achieve good cache
    performance.  Lens is also able to do batch parallel training
    on multiple processors.

-EFFICIENCY: Lens makes efficient use of memory, so it can handle
    larger networks and larger example sets.

-FLEXIBILITY: Because it includes a scripting language, the user
    has considerable flexibility in writing procedures to
    automatically build, train, or test customized networks.
    Unit input and output procedures can be composed from a
    variety of sub-procedures, allowing many group types to be
    created without the need to add additional code.

-CUSTOMIZABILITY: Lens was designed around the principle that no
    simulator can satisfy all users out of the box.
    Sophisticated users will inevitably need to customize the
    simulator at the source code level.  But altered code always
    causes problems if a new generic version of the source is
    released.  Therefore, Lens includes facilities for easily
    extending network structures and creating new types,
    algorithms, and commands without altering the main code base.

You can find additional information about downloading and using
Lens at its homepage:

                 http://www.cs.cmu.edu/~dr/Lens


AVAILABILITY 
Lens is available free-of-charge to those conducting research at
academic or non-profit institutions.  Other users should contact
me for licensing information at dr+lens at cs.cmu.edu.


CAVEATS 
Lens is designed for simulating connectionist networks.  It is
not appropriate for compartmental membrane-level or differential
equation-based modeling.