Space and Precision

Abstract

So far, we have considered neural networks with two types of resource constraints: time, and the Kolmogorov complexity of the weights. Here, we consider rational-weight neural networks in which a bound is set on the precision available for the neurons. The issue of precision comes up when simulating a neural network on a digital computer. Any implementation of real arithmetic in hardware will handle “reals” of limited precision, seldom larger than 64 bits. When more precision is necessary, one must resort to a software implementation of real arithmetic (sometimes provided by the compiler), and even in this case a physical limitation on the length of the mantissa of each state of a neural network under simulation is imposed by the amount of available memory. This observation suggests that some connection can be established between the space requirements needed to solve a problem and the precision required by the activations of the neural networks that solve it.