TRACE

1
jTRACE a reimplementation of the TRACE model of
spoken word recognition Ted J. Strauss, James S.
Magnuson, Harlan D. Harris

• TRACE
• The TRACE model (McClelland Elman, 1986) is the
  implemented psychological model of speech and
  spoken word processing with the broadest and
  deepest coverage of empirical data (see How
  TRACE works, below)
• TRACE is still used productively for theory
  testing, and is often discussed in papers on
  spoken word recognition
• Complex simulations often diverge from
  expectations
• But few people actually do simulations --
  requires significant technical skill
• Challenging even for experts to extract
  analyze output, do batches, or extend

• Why jTRACE?
• A user-friendly, platform-independent
  implementation would encourage more simulation,
  less speculation
• An open-source, powerful and extensible version
  would promote active development of the model
• We have developed jTRACE to fit these needs
• Easy to use -- GUI-based, no programming required
• Platform independent Unix, Macintosh, Windows --
  anything with Java
• Powerful GUI- or XML-based scripting facilitates
  large batches of simulations
• Extended includes graphing and analysis tools
• Extensible Conforms to contemporary programming
  standards, well-commented, open-source. Freely
  available at
• http//maglab.psy.uconn.edu/jtrace/

• jTRACE GUI
• Graphical user interface (GUI)
• Easy access to parameters (can reset to defaults)
  
• Simulations are visualized as animations
• Multiple representations -- matrix, floating
  unit

• Scripting
• Scripting automates preparation, running and
  analysis of simulations.
• Included templates make scripting easy (users can
  add more)
• Decision rules scripting greatly simplifies the
  analysis of large groups of simulations
• Scripting is done using a graphical interface, so
  no code to learn (but can be done directly in
  XML)
• Scripts can be run on a remotely (via ssh or
  telnet) without GUI
• Can be simple (single simulation) or complex --
  running multiple simulations on multiple (or all)
  words in a lexicon
• Examples of scripting elements
• Iteration over lexical items
• Iteration over parameters
• Conditionals
• Decision rules
• Functions
• Saving
• Graphing

• Other new features in jTRACE
• Ability to save/export/load simulations with
  varying levels of detail
• Comes with a gallery of classic simulations
  (user can add to this set)
• Makes sharing and replication easy
• New parameters
• Frequency (3 implementations from Dahan et al.,
  2001)
• Input and unit noise (stochasticity)
• Flexible creation of ambiguous phonemes and
  continua
• Luce choice rule, useful variants
• Validation tool for comparing jTRACE with
  original C implementation
• Source code
• Freely available, conforms to W3C standards for
  extensibility
• Object-oriented
• Commented and documented

• Graphing features Time course plots of
  activations or response probabilities

• Flexible selection of items to include, units to
  select Can be generated automatically via
  scripting
• Graphs can be saved to standard formats

• Validation and replication
• Compared jTRACE with cTRACE (original
  implementation)
• jTRACE is not a line-by-line reimplementation
• cTRACE uses pointer functions not available in
  Java
• jTRACE is object-oriented, among other
  improvements
• Compared cTRACE and jTRACE on 15 prior simulation
  projects (250 individual simulations)
• Metric Scaled Mean Absoulte Difference (SMAD)
• v 4-dimensional array of C and Java simulation
  data
• Min and max are bounds on the activation unit
  values
• A simulation data-set consists of about one
  million cells
• Difference between jTRACE simulations and
  original TRACE simulations is always less than
  0.007.
• Conclusion jTRACE is a faithful reimplementation
  of the TRACE model.

• Applications
• Research
• Easy to do TRACE simulations -- novice modelers
  can be doing their own simulations within an hour
• jTRACE is flexible enough to extend to new
  paradigms, parameters, etc.
• Scripting permits exploration of TRACEs
  parameter space
• Education
• Course demonstrations
• User-friendly enough to use for course labs

How TRACE works

• Interactive-activation network with dynamical
  properties, graded activation, decay
• Pseudo-spectral acoustic-phonetic input applied
  to feature layer in time slices corresponding to
  about 10 msecs (user can enter phonemic strings)
• Feedforward connections features to phonemes,
  phonemes to words

• Feedback connections words to phonemes (phoneme
  to feature off by default)
• Lateral inhibition between units within layers
• Each layer consists of multiple copies of each
  unit, aligned at different points in time (solves
  the segmentation problem by brute force)