Representing a Neural Network in Hadoop
A neural network is a graph structure
From computer science perspective, a neural network is a graph structure.
One common way to represent a graph is via an adjacency matrix. As we
have seen the below figure in the Neural Model,
the neural network is represented as a N*N square matrix, where a
cell value Mij indicates and edge from vertex ni
to vertex nj.
Using matrix is straightforward, but the matrix we are seeing is sparse, that is, a lot of cells are empty, and it is not algorithmically efficient to use a matrix representation to process the neural network.
For computer scientists, it is more natural to represent a graph with adjacency lists, in which a vertex is associated with neighbors that can be reached via outgoing edges.
In our implementation, the graph structure for a specific neuron is defined as:
[meta-data associated with the neuron][adjacency list]
meta-data associated with the neuron follows the schema below:
neuron ID, neuron type, time step, parameters, recovery variable,
membrane potential, input synaptic currents, firing flag, channel
adjacency list follows the schema below:
neuron ID-1:synaptic weight-1, neuron ID-2:synaptic weight-2,
... neuron ID-n:synaptic weight-n
In Hadoop, we used a custom writable NeuronWritable to store
meta-data associated with the neuron and Hadoop’s own Text format
for adjacency list.