Fig. 1: Illustration of the proposed HNN framework.
From: A framework for the general design and computation of hybrid neural networks
a Homogeneous information flow includes synchronous-continuous signals (solid orange lines) of ANN neurons (orange circle) and asynchronous-discrete signals (dashed green lines) of SNN neurons (green circle), respectively. Each case includes direct transmission (left, sharp arrow head) and indirect modulation (right, square arrow head pointing to parameters θ of modulated neurons). b Hybrid information flow is transformed by HUs (blue squares). c HUs have several basic computation steps, including truncating (W(t)), filtering (H(t)), non-linearity (F), and discretization (Q). These operations can be achieved by knowledge-driven manual design or data-driven automatic learning. d Designable HUs are configured according to the target coding schemes by prior knowledge and predefined mapping (The long time-scale red window for rate coding, short time-scale yellow window for timing code (e.g., synchrony)). e Learnable HUs can be configured in three learning ways: (1) jointly training with frontend/backend networks, (2) independent training, and (3) training with complete models.
