Spiking neural networks carry a compelling promise (brain-like computation at a fraction of the energy), but the gap between promise and practice remains real. This page sets out the main obstacles honestly, including the places where the field's headline claims deserve scrutiny, and the directions most likely to close the gap.
Training Difficulty
The root difficulty is the one that motivates the learning approaches: the spike is non-differentiable, so ordinary backpropagation does not apply. The dominant workaround, the surrogate gradient (Neftci, Mostafa & Zenke, 2019), makes training tractable but not free. Deep and recurrent networks are trained by backpropagation-through-time, whose memory cost grows with the number of timesteps and which inherits the vanishing- and exploding-gradient pathologies of recurrent networks, now compounded by the discrete spike. Performance is also sensitive to parameters a conventional network never exposes (the timestep count, membrane time constants, the firing threshold, and the shape of the surrogate), which is why Eshraghian et al. (2023) is framed as a tutorial for navigating them. The honest summary is that training is tractable but costlier and fussier than for a standard network.
The Accuracy Gap
Historically, spiking networks trailed deep conventional networks on large benchmarks such as ImageNet, closing the gap only by using many timesteps or by borrowing architectures wholesale from the conventional world (Roy, Jaiswal & Panda, 2019). That gap has narrowed sharply with spiking transformers: Spikformer (Zhou et al., 2022) reported around 74.8% top-1 on ImageNet at just four timesteps, though a residual gap to the best same-scale conventional models generally remains. Underlying it is a three-way tension between latency, accuracy and energy: adding timesteps tends to buy accuracy at the cost of the very efficiency that motivates the approach.
Hardware and Software Co-design
The efficiency benefits of spiking networks do not come from algorithms in isolation; they emerge only when sparse, event-driven computation is matched to event-driven hardware, which is why Roy, Jaiswal & Panda (2019) frame progress as a problem of algorithm–hardware co-design. Yet the software ecosystem is young and fragmented next to the mature toolchains of conventional deep learning, and each neuromorphic chip exposes its own programming model with limited portability. The Loihi survey of Davies et al. (2021) is candid that only recently have neuromorphic chips produced rigorously competitive, quantified results.
Benchmarking and Reproducibility
For much of its history the field lacked standardised, fair benchmarks, making genuine progress hard to measure and cross-paper comparisons unreliable, the explicit motivation for NeuroBench (Yik et al., 2025), a community framework offering common metrics in both hardware-independent and hardware-dependent settings. Energy comparisons are a particular hazard: Yan, Bai & Wong (2024) show that prevailing evaluations often count only arithmetic operations while ignoring data movement and memory access, an omission that can reverse the conclusion.
The Energy Question
This deserves stating plainly, because it is the field's most over-claimed point. The energy advantage of spiking networks is real primarily on event-driven neuromorphic hardware with sparse activity; it is not automatic. On a GPU, whose libraries are optimised for dense arithmetic, a spiking network can be slower and costlier than an equivalent conventional network. Yan, Bai & Wong (2024) quantify the crossover: under typical conditions a spiking network must keep its average spike rate low (on the order of a few per cent of neurons active) before it beats an equivalent quantised network, because event-driven weight reuse can require accessing each weight many times over the timesteps. The advantage is genuine in the right regime; the error is to assume it holds everywhere.
Future Directions
Several threads are converging. On learning, e-prop (Bellec et al., 2020) offers a mathematically grounded approximation to backpropagation-through-time that runs forward in time, opening the door to biologically plausible, on-chip training. On hardware, Loihi 2 and its successors aim to deliver quantitatively competitive results with richer neuron models and native on-chip plasticity. Event-based sensors continue to mature as a natural front-end that aligns sparse sensing with sparse computation. And the architecture is scaling: spiking transformers and generative spiking language models such as SpikeGPT (Zhu et al., 2023) are pushing toward regimes once thought out of reach, even if frontier-scale competitiveness remains unproven. The strongest near-term case for spiking networks is not to replace deep learning wholesale, but to own the latency- and energy-constrained, event-driven workloads where conventional networks are a poor fit.