SNN-LPCG: Spiking Neural Networks with Local Plasticity Context Gating for Lifelong Learning

1. The experiments focus on a somewhat simplified data setting. It would be convincing if more experimental results on more continual learning benchmarks are available.
2. Lack of satisfactory explanation of how “sluggish” neurons benefit continual learning despite sub-optimal final performance. While the paper demonstrates the effectiveness of sluggish neurons in capturing human cognitive behavioral patterns in context-dependent decision-making tasks, it does not provide a clear interpretation of how these neurons affect the overall performance of the model.
3. Distinct local plastic learning rules are applied in single- and multi-spike SNN-LPCG, respectively. Such comparison would be somewhat unfair if no further justifications are provided.

1. The related works lack references to continual learning with SNNs in the recent year.

2. The presentation of the parameter setting is not sufficiently clear and detailed. Please clarify it.

3. This paper lacks a crucial illustration that highlights the significance of cognitive control in the context of solving lifelong learning tasks. It would greatly enhance the paper's clarity to include a compelling explanation of why cognitive control is necessary in addressing these tasks.

• I had substantial difficulty in reading and understanding this work. I am sure this is in part due to a lack of expertise in SNNs but I do believe that overall the presentation is poor. See questions/clarifications below
• It is still not clear to me how context-dependent gating - a fundemental feature of this work - is actually implemented in this work. So far as I can tell, the local plasticity rules (STDP in eq 2, Oja in eq 7) provide the implementation for context-dependent gating (one subsection is called 'context gates by STDP'). How these rules are context dependent, or provide context gating, I don't understand. I believe Figure 1 provides the best illustration of how this is employed, and there is some text context input is used in 3.1, but I am desperate for a more formal description/equations (where is the context signal in eq 2/7? how do these mediate synapses in the SNN?). Moreover, if these are indeed the means for context-dependent gating, the fact that the single-spike and multi-spike models employ different rules (STDP + Oja respectively) makes it very difficult to identify the key computation which enables lifelong learning. Or should I understand that many types of local plasticity rules are (perhaps in their own way) beneficial for learning multiple tasks?
• Relatedly, it is not clear to me why the proposed model performs better than vanilla SNNs for remembering previously learned tasks. Can the authors suggests what latent factors are alleviating catastrophic forgetting? Most figures are dedicated to accuracy, selectivity and choice metrics, which are certainly helpful to see, but I would like to better understand why/how these local plasticity rules seem to preserve important weights and maintain performance. The authors do consider weight magnitude of STDP vs vanilla networks, but I don't see larger weights by themselves are beneficial for catastrophic forgettnig (and as the authors describe, these are prone to instability)
• It is unclear to me how the proposed model extends upon the Flesch et al. (2022) model. Many of their results seem very similar and capture human data well. What is the Flesch model lacking that this model possesses? The authors argue that the "reference network exhibits poor performance on the 1st task-selective neuron’s activity. This phenomenon indicates that our model has better biological plausibility and neural dynamics during human cognitive control." I don't understand this. What does poor performance on activity mean? Both models seem to learn equally well (if anything the reference model a tiny bit better). Is low selectivity less biologically plausible?
• A more detailed description of the task is necessary somewhere in the main text.

1. Overall the manuscript is not of publication quality, as it is not well-written and contains many errors, oversights, typos, figure organization problems, and a general lack of clarity. Some examples of these problems are listed below. While there are some promising results, I believe the manuscript needs a major rewrite and careful re-reading before resubmitting to another conference or journal.

2. Sluggish neurons appear to be core to the authors’ model, yet I was unable to find them defined. In the model, $\alpha$—which is used to define the sluggishness in the figures as ($1-\alpha$)—refers to the numerator of the surrogate gradient function $g(x)$ and is said to be set to 2.0 (note, $g(\cdot)$ is doubly defined, as it refers to the Heaviside function in Eq. 1).

3. The authors observe that their method results in more task-selective neurons compared to the “Reference” Flesch et al. 2022 model, and that this “indicates that our model has better biological plausibility and neural dynamics during human cognitive control.” But as no biological evidence from humans is provided for task-selective neurons, nor their relative proportion, no biologically-based adjudication can be made between the 2 models from this criterion.

   • Moreover, Fig. 3B top right is only brought up with the discussion of Fig. 4b. These panels should appear together and in order of reference of the main text

4. Given that the authors are aiming for greater biological plausibility in the context of catastrophic forgetting, the model neuron used should minimally be the standard LIF (leaky integrate-and-fire) neuron, not one that does not have a leak term.

5. The main Flesch et al. 2018 experiment should be briefly described in the Introduction, not in the Supplemental Material, so that the reader can better understand the motivation behind the model and the analysis thereof without having to download a separate file

   • For example, while “contextual signals” is mentioned and later modeled, the reader can’t know what that means with respect to this experiment without an initial explanation of the experimental setup and task(s)
   • Similarly, the fact that subjects received a reward is suddenly brought up when introducing the loss function rather than at the beginning

6. The authors compare their model to different models at different times. E.g., they compare their model to OWL / XdG in Fig. 3a, and then to the “Reference” model from Flesh et al. 2022 in Figs 3 and Supp. Fig. 1.

7. The multi-spike model is not well described, with many variables shown but not described. E.g., $I(t)$ is mentioned but does not appear in the model, it’s unclear what $n$ is, and while several of the variables will be easily inferred, each should be well-described after appearing in the model

8. Generally, there seem to be too few trials to make firm conclusions about the models’ final accuracies, as many have downward trends at the end of the trials that disallow any final conclusions. Examples include Fig. 2a, STDP and Multiple STDP and Fig. 3a OWM and XdG.

9. There are a number of models, analyses, and other important aspects that are not well described. For example:

   • XdG, plotted in Fig. 3a, is not described or referenced (it comes from Masse et al. 2018) in either the main text or the figure caption
   • The “Reference” Flesch et al. 2022 model is not described.
   • The choice matrices are not explained at all, and the linear model is only cursorily described (“the two features are learned chaotically and ignore the context cue” in Fig. 3 caption).
   • Statistical significance tests in Fig. 3a are not described at all
   • Error bars/shading not described at all

10. Fig. 3 needs to be split into 2 separate figures, and more panels labeled (A/B/C, i/ii/ii) to avoid top/bottom etc. labels. Currently the caption is difficult to read and associate the text to the panel under discussion.

11. I could not find any description or explanation in the main text for Fig. 3b bottom (Choice Fraction).

Minor

12. The second “contribution” seems to simply provide some detail to how the framework proposed in contribution 1 is implemented rather than to describe a contribution to the field.

13. Fig. 1: Top and bottom mixed up between figure and caption

14. It’s indicated that the spike times are mapped onto the z-domain, but no explanation of the z-domain is given, nor is it obvious as to why exponentiating the spike times would simplify the implementation

15. Some of the main text for Fig. 2A reads “The accuracy of the first task finally stays at about ∼75%.” Yet there appears to be a clear downward trend, rather than plateauing, at the end. It seems this comment should be taken out without further trials.

16. Fig. 2a needs to have a legend.