Modifying the DeSTIN Perception Architecture to Enable Representationally Transparent Deep Learning

Bridging the gap between symbolic and subsymbolic representations is a – perhaps the – key obstacle along the path from the present state of AI technology to human-level artificial general intelligence. A companion paper to this one describes a novel approach to achieiving this bridging via incorporation of a subsymbolic system and a symbolic system into a integrative cognitive architecture. This paper describes a key ingredient of this hybridization, which is also interesting in its own right: a modification of the DeSTIN deep learning based perception system, in a way that renders it ”representationally transparent,” meaning that when different parts of the deep learning network represent similar patterns (with similarity defined via affine transformations), this is immediately apparent via inspection of the state of the network. With DeSTIN as a case in point, it is argued that representational transparency is a desirable property for deep learning systems to have, for integration with other AI components as well as for other reasons, and that this can viably be achieved without substantially sacrificing their