Two different paradigms， symbolism and connectionism， have evolved from contemporary cognitive research. The computational?representation of symbolism is a "syntactic image" based on the language?of?thought hypothesis， which is featured by the separation between content and its carriers and by the context?independence of symbols. Deep learning has been an innovation and depth to connectionism technique， whose architectures are characterized as a kind of sub?symbol computation with such features as distributed processing and superposition storage， context?sensitivity， and prototype extraction， has demonstrated a series of anti?representational features which have been reflected by operations in deep networks that do not target explicit conceptual representations， and has driven the rise of anti?representationism in cognitive philosophy. Exploring unity between symbolism and deep learning architectures in some degree based on fully understanding their representational approaches may be a worthwhile goal.