In Search of Connections in Weightless Neural Networks

This work is based on a logical neuron model without weights, the Random Access Memory [1]. For the reason of memory consumption, RAMs cannot be completely connected in a bigger network, but the structure of connectivity is decomposed into n--tuple connections which causes additional properties of generalization [2]. Here a learning procedure for a feed-forward RAM network is presented which allows the neurons to exchange their input connections to resolve target conflicts in their function and thereby optimize internal representations. 1 Reduction of RAM Sizes in Layers of Neurons A RAM neuron can be trained by merely entering the desired responses into the memory locations addressed by its input (see RAM in Fig. 3). But difficulties arise with conflicting targets for the same event, which can be resolved by using larger RAMs or combining the outputs of several neurons in a layered network architecture, e.g. PYRAMID [1]. When neuron layers are used, internal states need to be define..