Fig. 4. HRN for VQA task.

a Two ANN-based parsers extract vision and language features, and an SNN-based analyzer executes a two-phase process to construct a graph and reason with it. In the construction phase, the SNN is initialized with prior knowledge, and then the independently trained learnable HUs transmit visual information into spike trains, which activate sensory neurons and intermediate neurons simultaneously in the SNN to embed external information according to the Hebbian rule and one-shot plasticity. In the reasoning phase, designable HUs transmit language information into polychromous activations⁵² activating command neurons to execute the reasoning process. b The accuracy of the CLEVRER validation set according to different task types. The reasoning results are compared with CNN-LSTM³⁵, MAC(V+)³⁶, NS-DR³⁷, DCL³⁸, Aloe³⁹, and VRDP⁴⁰, respectively. c The plot of calculation latency and the number of spiking neurons against the increase of target objects. d Reasoning robustness to the frontend anomaly data.