Changes 4.0.7
From Emergent
Contents |
Major new features and fixes for Emergent release 4.0.7
Analysis/Display Tools
- Network display can show weights into/out of a selected unit as lines projecting to other units -- this is more "naturalistic" and also can be useful when viewing activations on the units, to assess the fit between activations and weights. Control is in the netview control panel, including a threshold for only showing the strongest weights.
- In addition, you can project weights for the selected unit across multiple layers (and view the results as lines as well as on the units) -- useful for analyzing the coding of deeper hidden units relative to input and output patterns.
- Activation-based receptive field computation (was a stat in 3.2) is now available -- also useful for understanding what hidden units are coding for, by a "reverse correlation" technique of weighted averages on unit activations throughout the network. Available as ActBasedRF object and ActBasedRField.prog program in program library.
Misc New Features
- taCanvas for creating images using basic drawing operations supported by Qt QPainter -- useful for programmatically creating visual test stimuli (see demo/data_proc/canvas_drawing.proj).
- LeabraStartup program and STARTUP_RUN flag on programs -- replaces separate startup script for running projects in the background.
- Added WtInitPrompt program element to replace usage of IfGuiPrompt in prompting whether to initialize the weights in the train program. This checks to see if the network has already been trained before prompting -- cuts down on the number of prompts. See the link for how to update.
Minor Fixes etc
- Removed lots of spacious frames etc to maximize content display in various windows.
- Reverted Mac packages to Qt 4.2.3 to avoid serious drag-and-drop and other layout bugs with 4.3.1.
- The 2x2 project view configuration is now supported.
- Added new demo projects in demo/data_proc, including data_calc_loop.proj (illustrating how to use the powerful DataCalcLoop program elements) and epoch_log_analysis.proj (handy tools for analyzing training logs from networks).
- Numerous bugs fixed etc!
