A map of the ANN→SNN gap: discrete spikes, temporal coding, local plasticity, surrogate gradients,
and conversion pipelines. This page treats the field as competing objective functions, not one score.
arXiv 1804.08150Core review Tavanaei et al. 2018/2019Episode lens engineering vs biology vs benchmarksExtra IDs found in transcriptnone beyond 1804.08150
Method Families
3
STDP/local learning, direct training, ANN→SNN conversion
Core Tension
4 Axes
accuracy, latency, energy, plausibility
Episode Claim Under Inspection
Progress is strongest where objectives are constrained: conversion for accuracy, direct training for native supervision, STDP for locality.
Landscape: One Field, Three Objective Functions
Each method family optimizes a different compromise. Hover the heatmap, then compare how each family routes around the non-differentiable spike.
Objective Heatmap
weak fit mixed fit strong fit
Training Route Diagram
Direct training bends gradients around the spike. Conversion preserves ANN competence. STDP keeps updates local but usually gives up end-to-end optimization.
Dynamics: What Information Looks Like in a Spiking System
Switch between rate and temporal coding. The same stimulus can mean “how often” or “exactly when,” which changes how a learner should assign credit.
Spike Raster and Membrane State
STDP Learning Window
Positive Δt means pre-synaptic spike arrives before post-synaptic spike, strengthening the synapse. Reverse the order and the weight depresses.
Tradeoffs: Accuracy Is Not Energy Is Not Latency
Toggle the scoreboard. The visual point is not the exact numbers, but how rankings change once you judge by a different metric.
Method Comparison
Mock values reflect the episode’s qualitative takeaway: conversion is accuracy-heavy, direct training is balanced but harder, STDP is local and sparse yet less competitive on standard static benchmarks.
Latency vs Energy Frontier
Changing timestep budget reshapes the cost surface. “Efficient” SNN claims depend on simulation target, hardware, coding scheme, and allowed settling time.
Timeline: How the Taxonomy Expanded
Step through key phases. The field starts with computational arguments, then pushes conversion and surrogate gradients, then expands into transformer and state-space variants.
Research Arc
Architecture Pressure Map
Rows are research themes; columns are epochs. Later work broadens the design space, but also makes cross-paper comparisons less clean.
References
Deep Learning in Spiking Neural Networks — Tavanaei et al. (2018)
arXiv:1804.08150
Networks of Spiking Neurons: The Third Generation of Neural Network Models — Maass (1997)
Scholar link
Training Deep Spiking Neural Networks Using Backpropagation — Lee et al. (2016)
Scholar link
Conversion of Continuous-Valued Deep Networks to Efficient Event-Driven Networks — Rueckauer et al. (2017)
Scholar link
Fast-Classifying, High-Accuracy Spiking Deep Networks Through Weight and Threshold Balancing — Diehl et al. (2015)
Scholar link
Surrogate Gradient Learning in Spiking Neural Networks — Neftci et al. (2019)
Scholar link
SLAYER: Spike Layer Error Reassignment in Time — Shrestha & Orchard (2018)
Scholar link
AI Post Transformers: Directly Trained Spiking DQNs for Atari
Episode audio