Fractal connection structure: effect on generalization in supervised feed-forward networks

Fractal connection structure within the layers of a multilayered feedforward net has been studied in this paper. Fractal connection structure ensures modularity, easy hardware implementation and resembles biological neural system more closely than fully connected layered architecture. Simulation on sonar signal for underwater target classification problem shows that fractal net with fractal dimension around .9 with average connectivity 80% performs better than the fully connected net of same size (same number of neurons) in terms of classification accuracy and generalization behaviour to unseen samples