<title>Negative transfer problem in neural networks</title>

Abstract

Harlow, 1949, observed that when human subjects were trained to perform simple discrimination tasks over a sequence of successive training sessions (trials), their performance improved as a function of the successive sessions. Harlow called this phenomena "learning-to-learn" . The subjects acquired knowledge and improved their ability to learn in future training sessions. It seems that previous training sessions contribute positively to the current one. , observed that when a neural network (using the back-propagation model) is trained over successive sessions, the performance and learning ability of the network degrade as a function of the training sessions. In some cases this leads to a complete paralysis of the network. Abunawass & Maki called this phenomena the "negative transfer" problem, since previous training sessions contribute negatively to the current one. The effect of the negative transfer problem is in clear contradiction to that reported by Harlow on human subjects. Since the ability to model human cognition and learning is one of the most important goals (and claims) of neural networks, the negative transfer problem represents a clear limitation to this ability. This paper describes a new neural network sequential learning model known as Adaptive Memory Consolidation. In this model the network uses its past learning experience to enhance its future learning ability. Adaptive Memory Consolidation has led to the elimination and reversal of the effect of the negative transfer problem. Thus producing a "positive transfer" effect similar to Harlow's learning-to-learn phenomena. BACKGROUND Artificial Neural NetworksArtificial neural networks (also known as connectionist models, adaptive neural networks, neural networks, adaptive systems, neuromorphic systems and parallel distributed processing) have a rich history, numerous successes and a wide appeal. The appeal of artificial neural networks includes the potential of brain-like computations and human-like learning and intelligence. We believe that for an artificial model to be a viable model of learning and intelligence, the model must be capable of matching the experimental results of its natural counterpart. Loosely stated, the artificial model must be able to mimic and exhibit the behavior of the natural one. Artificial neural networks have problems that limit their ability to mimic the behavior of the natural ones; one such problem is the topic of this paper. In the paper we are limiting our focus to the back-propagation (BP) model [8].Artificial neural networks have a similar topography to directed graphs (see Figure 1). The nodes in the graph are known as units or neurons. The edges of the graph are known as connections or synapses. The units are linked to each other via connections. The connections are unidirectional. Each unit has output and input values. Each connection has a weight value. The weights of the connections represent the long term memory (LTM) of the network. LTM contains the learning the network has accomplished. LTM is the perma...