Update 3.4 Released
May 27th, 2006
We have rolled out a new update, (1.0)/3.4 which contains a number of bug fixes and some new features. The update will be automatically downloaded and installed when you start Synapse (provided that you are connected to the internet)
New features:
- The new Updater2 update engine with an improved versioning system, download system and support for a more advanced release structure.
- Improved CSV input format that provides more features for handling string data (enumeration and removal in addition to the existing expansion):
- Generalized Hebbian Algorithm (GHA) update rule. A new Hebbian type update rule that performs PCA. The difference between it and Oja’s rule is that GHA results in sorted principal components.
- Splash screen, minor improvement with verbose loading of plugins
- Start screen, some visual improvements
- Two more plot line styles and marker support added to the Scatter Plot:
- Data embedded in solution is now compressed
Bug fixes:
- Component GUI flicker fixed
- Select Filter inadvertent sort fixed
- Hebbian component GUI refresh in training fixed
- Script Filter, disconnected GUI elements removed
- Unstage (scaling) in the Sink component fixed
- Column resize in GridView enabled
Good work, the csv file format is much bettern now. Thanks.
I had some trouble installing it until I realized that my firewall was blocking it out. Doh!
Hi
I am wondering if you use any other second order error minimization methods than QuickPRop. I mean Levenberg Marquardt, Conjugate Gradient, Newton method etc. in MLP.
Hi
I am also wondering if you do now or plan to use sensitivity approaches to network pruning in MLP to optimize the networks as this approach may be more useful in deleting links more meaningfully and methodically?
Hi,
You can use Levenberg-Marquardt as well with components that support gradient optimization. As for network pruning, that’s what the optimizers are for (GA, particle swarm etc). There are also implicit pruning methods through unsupervised training (weight decay…).
As for sensitivity analysis, while you can use it for pruning purposes, it is something that I would not recommend. The sensitivity approach is simply a too blunt instrument as it doesn’t take into consideration the complexities of the non-linearity of the systems and the correlation of data. In most cases it will do more damage than it will help.
The best pruning approach is through the use of global optimizers that optimize the whole system topology.
–Thomas