- Rating
- 2
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes a minimal astrocyte-inspired computational module for spiking neural networks (SNNs). The module integrates local spike activity over space and time, generates events when the accumulated activity crosses a threshold, and uses a persistent modulatory state to regulate synaptic transmission and plasticity. The authors derive a mean-field threshold-crossing law and a critical input rate, and analyze the trade-off between noise reduction and spatial bias.
The method is evaluated on an unsupervised Diehl–Cook SNN using MNIST, as well as pretrained SNNs on N-MNIST and DVS128-Gesture. Under 75% dynamic salt-and-pepper corruption on MNIST, the proposed model achieves 78.1% accuracy compared with 9.8% for the unmodified baseline. Additional experiments examine robustness to injected event noise, adaptation to changing inputs, and the limitations of the proposed mechanism under structured noise.
Critical strengths
1. Substantial robustness improvements under controlled noise
The proposed method achieves a large improvement on heavily corrupted MNIST, from 9.80% to 78.06% accuracy at , with consistent results across four random seeds. The improvement persists across several corruption levels. The ablations in Table 4 also show that both input-transmission and plasticity-rate modulation contribute to performance. These results provide evidence that the proposed mechanism can be effective under the tested conditions.
2. Analysis of the operating regime and failure conditions
The paper goes beyond reporting classification accuracy by deriving a critical input rate and examining when the proposed detector is expected to work or fail. The observed performance collapse occurs near the predicted loss of mean-field rate separation. The authors also evaluate native event streams and report negative results under structured noise. These analyses help characterize the operating conditions and limitations of the method, rather than presenting it as uniformly beneficial.
Critical weaknesses
1. Limited novelty and unclear computational contribution
The proposed method combines spatial pooling, leaky temporal integration, thresholding, and local gain modulation (Eqs. 1–4). These are established computational operations. Previous work has already modeled astrocytes as spatiotemporal integrators (Gordleeva et al., 2019) and incorporated astrocytes into SNNs for image recognition (Lorenzo et al., 2025).
The theoretical results mainly characterize the threshold-crossing behavior of a leaky integrator and the variance reduction from averaging independent Poisson inputs (Eqs. 5, 9–10). It is unclear what new computational principle this formulation establishes, or what capability it provides beyond existing adaptive filtering and gating mechanisms.
2. The main experiment is closely matched to the proposed mechanism
The reported improvement from 9.8% to 78.1% under 75% MNIST corruption is substantial, but the experimental setup is particularly favorable to the proposed detector.
Under the corruption model in Eq. 12, the extreme foreground and background firing rates become 37.5 Hz and 22.5 Hz at . The detector is calibrated at this same corruption level, with a critical rate of approximately 31.9 Hz, positioned between these values. The experiment therefore creates a rate-separation problem that the proposed temporal integration and thresholding mechanism is specifically designed to solve.
Moreover, Table 2 shows that a simple hard local gate already achieves 75.61% accuracy, compared with 82.31% for the full astrocytic model. Although this is not a fully matched event-domain comparison, it suggests that conventional gating can recover much of the lost performance.
The paper does not adequately establish whether the full astrocytic mechanism provides a meaningful advantage over comparably tuned conventional gates.
3. Limited generality and uncertain practical value
The only end-to-end training experiments use the classical Diehl–Cook SNN on MNIST. The N-MNIST and DVS128-Gesture experiments insert the module into frozen pretrained networks. No larger, modern SNN architectures are evaluated.
The results also depend strongly on noise statistics. In Table 5, under hot-pixel corruption at , the original astrocytic model achieves 77.45%, below the unmodified baseline's 82.72%. Under temporal-burst corruption, it also underperforms (90.80% versus 92.54%). The alternative active-channel detector improves hot-pixel robustness but performs substantially worse under temporal bursts.
These results raise doubts about whether the method offers general robustness rather than benefits specific to selected noise conditions. The paper also lacks measured inference latency and energy consumption, leaving the practical cost of the additional states and event processing unclear.
Questions for authors
-
Why does the MNIST baseline collapse to nearly random accuracy? Figure 2(d) reports only 9.8% accuracy at (p=0.75), despite the baseline being trained on corrupted inputs. What causes this failure? Have the authors tried retuning the baseline's input gain, firing thresholds, and homeostatic parameters for the same noise conditions? How much of the reported improvement remains after doing so?
-
How much of the improvement comes from controlling spike activity? Table 3 shows that the unmodified network produces 77.99 spikes per image under corruption, compared with 22.46 for the proposed model at ( ). Could the authors compare against a conventional activity-normalization or adaptive-threshold mechanism that achieves a similar firing rate? This would help determine whether the improvement requires the proposed astrocytic dynamics.