Readr is a library using which programmers can create and compare neural networks capable of supervised pattern recognition without knowledge of machine learning. These networks are fuzzy-neuro systems with fuzzy controllers and tuners regulating learning parameters after each epoch to achieve faster convergence.
The learning algorithms provided by Readr are:
-
General Back Propagation: The standard neural network with no fuzzy controllers acting on it. It converges with the slowest speed within all networks provided by Readr. Also used as a benchmark for performance of other neural networks.
-
GBP with Momentum: A neural network with GBP and momentum applied to it. It is also used as a benchmark for other neural networks provided by Readr. About 200% faster convergence than GBP.
-
GBP with Fuzzy Gain: The network proposed by the paper we implemented. It passes the gain of the activation function to a fuzzy controller in order to maximize convergence speed. Approximately 1000% faster than GBP.
-
GBP with Fuzzy Momentum: The network fuzzifies momentum in order to enable fasster convergence. This is slower than GBP with Fuzzy Gain, but faster than the unfuzzified momentum.
-
GBP with Fuzzy Gain and Momentum: This neural network fuzzifies gain and momentum both in order to maximize convergence speed.
Note: All networks currently use log-sigmoid as their activation funcations and have 3 layers. These parameters will be made tunable in future releases.
To use Readr, import the library in your script as:
import reader as rd
The following functions allow you to deploy networks:
Returns a session object with a new neural network. This object is required for all operations on the network.
Trains the neural network with the image provided.
| Parameter | Description |
|---|---|
| sess | Session object for neural network |
| imagepath | Path to the image to train with |
| actualresult | Desired output |
| enableGainFuzzification | Fuzzy control for gain |
Returns: Convergence of the network. (~98%)
Reads the image and determines the pattern.
| Parameter | Description |
|---|---|
| sess | Session object for neural network |
| imagepath | Path to the image to get pattern from |
Returns: Pattern Matrix from output layer.
Reads the image and returns a numeric answer.
| Parameter | Description |
|---|---|
| sess | Session object for neural network |
| imagepath | Path to the image to get pattern from |
Returns: Node number at which output == 1 from the output layer.
provideMnistTraining(sess, numTrainingSamples, enableGainFuzzifization = True, enableEtaFuzzification = True, enableMomentumFuzzification=True, momentumConstant = 0.0)
Trains the network with MNIST dataset to up to 90% convergence. Current implementation depends on sensitivity, which is an internal variable. This tuning will be made automatic in future releases.
| Parameter | Description |
|---|---|
| sess | Session object for neural network |
| numTrainingSamples | Number of training samples to train the pattern with |
| enableGainFuzzification | Fuzzy control for gain |
| enableEtaFuzzification | Fuzzy control for learning rate |
| enableMomentumFuzzification | Fuzzy control for momentum (Does not work if momentum is disabled in the network) |
| momentumConstant | Enable or disable momentum in the network |
Returns: Convergence of the network. (~90%).
Resets the neural network.
| Parameter | Description |
|---|---|
| sess | Session object for neural network |
Saves the model in a file on the system. Useful to store networks once training is complete.
| Parameter | Description |
|---|---|
| sesison | Session object for neural network |
| filename | Path where the network should be stored |
Loads the model from a file on the system. Used when tarined networks are to be loaded into memory.
| Parameter | Description |
|---|---|
| sesison | Session object for neural network |
| filename | Path where the network should be stored |
All Readr source code is made available under the terms of the GNU Affero Public License (GNU AGPLv3).