A fast iterative learning strategy for Bi-directional Associative Memory

A fast iterative learning strategy for

Bi-directional Associative Memory

Hoa Thi Nong and The Duy Bui

Human Machine Interaction Laboratory

Vietnam National University, Hanoi

Email: [email protected]

Abstract. In this paper, we propose a new iterative learning strategy

of Bi-directional Associative Memory (BAM)to guarantee the recall of

all training pairs. In our learning strategy, pair weights in interactive

learning algorithm are modified so that learning process is faster and the

ability of recall is larger or equal to than other BAMs.

Keywords: Bi-directional Associative Memory, Learning Algorithm, Hop￾field Neural Networks.

1 Introduction

In this paper, we propose a new iterative learning strategy for BAMs that can

reduce significantly the learning time while keep the noise tolerance high. Our

strategy performs iterative learning until we obtain the condition that guarantees

the recall of all training pairs, meaning that our novel model converges in all

states. Updating connection weights is flexible by changing pair weights in an

iteration of learning process. As a result, speed of learning increases. Moreover,

we prove advantages of our novel model in theory and by experiments.

The rest of the paper is organized as follows. In section 2, we present our

novel learning strategy and prove advantages. Section 3 shows our experiments

and compares with other models .

2 Our approach

2.1 Our learning strategy

The goal of our learning strategy is to build a connection weight matrix that

assures our model always converges in all states. That means our model can

recall correctly all training pairs.

Assume that we want to learn N pattern pairs, (A1, B1),(A2, B2), ...,(AN , BN ).

Let qi, Ei be pair weights and energy of i

th pair, ε be the threshold of energy

to represent the stop condition of modifying pair weights, cond(i) be the lower

bounds of pair weight of i

th pair, and t is ratio of Ei. Our learning strategy

comprises of three steps can be described as follows.

- 182 -