Getting Started: Introduction

Lens is an efficient, yet flexible, neural network simulator that runs on a variety of platforms, is able to handle large, complex simulations, but is also reasonably easy for novices to operate. Lens has three main design objectives:

Lens is written primarily in C. The graphical interfaces and script interpreter were implemented using the Tcl/Tk toolkit version 8.3, involving a combination of Tcl scripts and direct C library commands. Lens runs on most Unix and Windows machines. There are not, at present, plans to port Lens to the Mac because of its various deficiencies.

The simulator currently provides five learning algorithms: steepest descent, momentum descent, "Doug's momentum descent", delta-bar-delta, and quick-prop, as well as a variety of unit and group input-combining and transfer functions. Flexible group types provide a wide array of unit behaviors. New functions can be easily added to extend the standard functions. Because delta-bar-delta and quick-prop require an extra field in the link structure, which can result in a noticeable performance degradation when those fields are unnecessary, special compile flags are used to produce version of Lens that support those algorithms.

Lens supports feed-forward, simple recurrent, recurrent backprop through time, and continuous RBPTT networks as well as deterministic Boltzmann machines and Kohonen networks. A variety of other backprop and non-backprop network types could be added with little trouble.

The Lens shell allows commands to be typed directly to a standard terminal or to a built-in graphical shell window and will interpret scripts written in the Tcl language, which offers more flexibility than most shells, including a variety of control structures and the ability to define new functions and new control structures. In addition to Tcl/Tk procedures, Lens provides over 100 commands for building, training, and analyzing networks.

Graphical interfaces provide convenient shortcuts to the most frequent tasks as well as access to the fields of the network data structures. These values of these fields can also be set and retrieved on the command-line or in scripts. The Unit Viewer displays the current training or testing set along with values associated with each unit, such as its activation or error derivatives. The layout of these units may be customized or generated automatically. The Link Viewer displays the connectivity matrix of the network, showing values associated with each link, such as its weight or error derivative. Finally, any real-valued variable in the network or the output of any command may be graphed over time.

Lens allows multiple networks to be loaded simultaneously. In many simulators, example sets are specific to a particular network. If the same set is to be used by multiple networks, it must be loaded multiple times. In Lens, on the other hand, example sets are independent of networks. Thus, an example set can be loaded once and used as both the the training and testing set of a network or shared by multiple networks. In addition, networks can be destroyed and rebuilt without the need to reload potentially very large example files.

Along with standard uni-processor training, Lens supports batch-level parallel training. Each processor contains its own copy of the network and example files. One processor acts as a server, farming out training batches to client processes. When a client completes a batch, it returns the accumulated link error derivatives to the server, which performs the weight updates and sends the client the new set of weights. This can substantially speed up training in some networks.

Douglas Rohde
Last modified: Wed Apr 19 18.39:58 EDT 2000