Snippet

From Piki

A snippet is a partial toplogy that has been saved for reuse. Snippets can be used to rapidly construct adaptive systems with one click.

Contents 1 Using snippets 2 Creating snippets 3 Synapse snippet library 3.1 Classifiers 3.1.1 Bayesian classifier 3.1.2 Support vector machine 3.2 Dynamic 3.2.1 Gamma One Layer 3.2.2 Gamma Two Layer 3.2.3 Gamma Recurrent Hybrid 3.2.4 Recurrent One Layer 3.2.5 Recurrent Two Layer 3.3 Expanders 3.3.1 Binary Expander 3.3.2 Tertiary Expander 3.4 Focused 3.4.1 Elman Network 3.4.2 Time Delay Neural Network 3.5 Specialized 3.5.1 RBF Network 3.5.2 Wavelet One Layer 3.5.3 Wavelet Compound 3.6 Static 3.6.1 MLP One Layer 3.6.2 MLP Two Layer 3.6.3 Generalized One Layer 3.6.4 Generalized Two Layer 3.6.5 Modular Network 3.7 Unsupervised 3.7.1 Anti-Hebbian Novelty Filter 3.7.2 Hebbian PCA 3.7.3 Projected Competitive

Using snippets

To insert an existing snippet right click on the work area (the white region in the middle of the screen). You'll get a pop-up menu from which you can choose "Insert Snippet". There you can select the snippet that you want:

Creating snippets

To save a topology as a snippet, first select all blocks you want to include (you can select all blocks by pressing CTRL+A). Right-click on any of the selected blocks to get a context menu:

Select "Create NetSnippet..". You will get a new window, in which you can write a name, author and description:

Click on Save, and you will be asked for a location. The default user snippet directory is under My Documents/Peltarion Synapse/Snippets.

Suppose we create a new directory called "Cattle" (under My Documents/Peltarion Synapse/Snippet) and save the file in it (CattleNet.synsnip).

Back in Design, right click on the work area, and select "Insert Snippet.." Notice now that htere is a "Cattle" catgeory with a "CattleNet" snippet.

Synapse snippet library

Synapse comes with a collection of snippets of popular types of systems.

Classifiers

Classifiers are used to categorize data into two or more classes. They are supervised algorithms.

Bayesian classifier

A simple and fast classifier that despite its simplicity is often capable of solving moderately complex classification problems. It uses the Bayesian block as the classification engine.

Support vector machine

A very powerful non-linear, binary classifier that can solve complex classification problems. It is computationally very slow and sensitive to changes in settings. It uses the support vector machine block as the classification engine.

Dynamic

Dynamic snippets use memory structures and are for processing time series. They require the dynamic XProp control system to run.

Gamma One Layer

Dynamic neural network of intermediate complexity suitable for time series modeling and dynamic filtering. The primary memory component is the gamma layer block.

Gamma Two Layer

Dynamic neural network of high complexity suitable for time series modeling and dynamic filtering.

Gamma Recurrent Hybrid

Dynamic neural network that combines gamma and infinite impulse response memories. Suitable for simple dynamic problems that contain multifrequency dynamics.

Recurrent One Layer

Simple recurrent dynamic neural network that incorporates infinite impulse response filtering. Solutions can become divergent.

Recurrent Two Layer

Recurrent dynamic neural network that incorporates infinite impulse response filtering. Relatively powerful for dynamic problems but solutions can easily become divergent.

Expanders

Expanders perform feature expansion (ie. transform categorical data from a single feature to one feature per category). The expanders use the fuzzy logic block for their operation.

Binary Expander

Expands one binary features to two binary nominal features. [0] -> [1 0] and [1] -> [0 1].

Tertiary Expander

Expands a three valued feature into three binary features.

Focused

Focused learning is a way of building systems that handle time series data while using a static control system. While not as powerful as true dynamic systems, they are considerably faster.

Elman Network

A dynamic neural network that can be trained with a static control system (faster).

Time Delay Neural Network

A focused neural network for solving temporal problems. Memory depth is controlled by the number of taps on the Gamma memory.

Specialized

More specialized types of networks.

RBF Network

A radial basis function neural network . It uses gaussian activation functions whose parameters are determined through unsupervised learning. It performs more localized adaptation than other types of neural nets.

Wavelet One Layer

A basic wavelet neural network (WNN). It can outperform a regular MLP on some function modeling problems. The basic operating principle is that it makes use of wavelets which are a specialized form of functions. The wavelets are scaled, transposed and rotated and combined together to make a better function fit.

Wavelet Compound

A compound wavelet network, that uses elements from both a WNN and MLP networks.

Static

Static systems are used for data that has no temporal dependencies.

MLP One Layer

The Multi-Layer Perceptron (MLP) is a basic static feedforward backpropagation neural network. It is a good starting point for most classification and function modeling problems.

MLP Two Layer

Static feedforward backpropagation neural network of intermediate complexity. With the right configuration and data, it is theoretically capable of solving any function modeling or classification problem.

Generalized One Layer

An extension of the standard One Layer MLP neural network that in many cases is capable of solving a problem more efficiently. It is suitable for function modeling and classification tasks where plenty of data is available.

Generalized Two Layer

An extension of the standard Two Layer MLP neural network that in many cases is capable of solving a problem more efficiently. It is suitable for function modeling and classification tasks where plenty of data is available.

Modular Network

A static neural network that has two main branches. Generally during adaptation the branches compete against each other, often resulting in a system that is capable of better generalization.

Unsupervised

Unsupervised learning means that no "correct" answer is provided as feedback to the system. Instead the algorithm has to make sense of the input data alone.

Anti-Hebbian Novelty Filter

A trained anti-hebbian novelty filter reacts to data different from what it has seen during training. This is essentially a Hebbian block with a GHA update rule with a negative learning rate.

Hebbian PCA

Uses a Hebbian layer block to perform Principal Component Analysis (PCA). If outputs is set to 1, it acts as a maximum eigenfilter.

Projected Competitive

Combination of two competitive components, one with projection output and the other with nearest unit output. Useful as a pre-stage to supervised networks.