Language Project
Neural Network

Neural networks have a mass appeal
 Simulates brain characteristics

Weighted connection to nodes
Neural Network cont.





A neural network consists of four main parts:
1. Processing units.
2. Weighted interconnections between the
various processing units which determine how
the activation of one unit leads to input for
another unit.
3. An activation rule which acts on the set of
input signals at a unit to produce a new output
signal, or activation.
4. Optionally, a learning rule that specifies how
to adjust the weights for a given input/output
pair.
Learning and predicting

Learning method is back propagation
 Error is calculated at result nodes
 This error is then fed backward through the net
 Adjustments to weight reduce error
 We look at our result nodes and Output node
which is the last node, and
 The difference we get is the error
Predicting

Predicting is summation of weights
 Dendrites have input value
 Input times weight gives value
 All value summed give total activation
Parameters

Internal parameters
 Learning rate
 Threshold

Dendrite initial value
 Value
 Rand

Activation formula
 Triangle
 Logistic
 linear
Elements

sense Gets input from the file
 Input is in sequence

dendrite Connects neurons
 Has weight
 Weight is a floating value from 0-1
Elements cont

soma –
 body of a neuron

synapse –
 It is used for connection. It determines
which dendrite will go to which neuron

result –
 results are supplied to this node
Parts of Language

The derived NN Engine consists of three
parts
1.
Framework initialization

Topology implementation
2.

3.
Link necessary files
NNL specific code
Processor

Process input using topology
Properties










Its imperative
Keywords are not case sensitive
Id’s are case sensitive
Last neuron of layer is one to one relationship
with result
We can implement two neural networks, take
the same sense values and get different
results to compare
Readable
Writable
We are using functions without declarations
Functions can take any values
Error will be caught in semantics
Input and Output

Network input
 Format in in in … in # out out …out
 Error checking for input file

Network output
 The output file is same format
 Populate empty output with predictions
Grammar Used













A-> O; A | D; A | N ; A | Y; A | S ; A | R;A| e
O-> soma I F --------body of a neuron
E->dendrite E
I-> id
Y->synapse II --------- a connection
F->function P
P->( P’)
P’->Z,P’| Z
D-> dendrite I F -------- input to neuron
N -> neuron E --- a neuron composed of soma and dendrite
S->sense Z I ---- information is supplied to this node
R-> result Z id ---- results are supplied to this node
Z->number
Sample Code

// create first neuron (Logistic and Triangle are
functions)
soma s1 Logistic(10, 2, 5);
dendrite d1 Value(1);
dendrite d2 Rand(1,2);
neuron n1 s1 d1 d2 ;
// create second neuron
soma s2 Triangle(3);
dendrite d3 Value(1);
dendrite d4 Rand(1,2);
neuron n2 s2 d3 d4 ;

// create second neuron
soma s3 Triangle(3);
dendrite d5 value(4);
dendrite d6 Rand(1,2);
neuron n3 s3 d5 d6;
// connect neurons
synapse n2 d2;
synapse n3 d1;
// input
sense 1 d1;
sense 2 d3;
sense 3 d4;
// out
result 1 n1;
Engine

Performs analysis on neurons
 Detects layer of neuron
 order neurons in a list by layer

Processes neuron
 Cascades to get predictions
 Performs back propagation

Input data
 Streams data from input file
 Check data for errors

Output data
 Writes results
Traversing neurons
Compiler

Lexical Analyzer
 Removes comments
 Conditions code
 Tokenizes

Parser
 Recursive descent
 Does not do static semantic
Compiler cont.

Semantics
 checks ids declarations
 Checks how ids are assembled

Code generation
 Transforms NNL code into java
 Adds the engine
System Interactions

Creating network

Using network
Example

Detecting letters
 detect L and T in a 5 by 5 grid
 Compare L and T – no false positive
 Compare random results (noise rejection)

Topology
 One layer network
 One output
 Twenty five inputs
Example cont.
Future Enhancements
Persistent data for each NNL program
 Easier training methods
 Keyword to generate redundant
declarations
 Visual connection tool
 Include parameter choices
 High level semantics

Questions