A Visualisation Tool for N-Dimensional Error Surfaces
Alexander Pudmenzky · 1998
Visualisation of N-dimensional surfaces is limited by the disability of the human observer to picture more than three dimensions. To overcome this problem lower dimensional slices of N-dimensional surfaces are often utilised. This method risks omitting important domains from view. A new technique is introduced here that allows the observer to visualise more than three dimensions without hiding any dimensions from view. Instead of omitting some dimensions completely, all dimensions are shown but only in the immediate neighbourhood of the present viewpoint. This is achieved by intersecting all 2-dimensional projections of each dimension using the present observer point as a common intersection point. The projections are spread in an either circular or horizontal stacking arrangement in 3-dimensions with the common viewpoint in the centre. We introduce this technique by comparing the weight spaces of two Artificial Neural Networks (ANNs) with different update algorithms during training an...