Representational drift under spontaneous activity - self-organized criticality enhances representational reliability

Reviewer #1 (Public review):

Summary:

The authors study criticality and drift in spontaneous activity observed in visual cortex of mice from existing data, and relate it to a model based on homeostatic plasticity. The main phenomena are power laws and an alignment across different neural representations that is maintained through drift.

Strengths:

The authors should be commended by making the effort of relating their model to experimental data. The mechanism that they propose has the advantage of being simple, and could unify various phenomena.

Weaknesses:

Introduction/abstract: General wording: the notion of reliability, which is key to the paper is not explicitly defined anywhere. The authors refer to some notion of information being preserved, but again, this is not clearly explained. A good example is the sentence "identical input signals exhibit significant variability but also share certain reliability across sessions". Depending on the definition of reliability, the sentence could be a contradiction. A similar issue appears when the authors talk about "restricted" representation. I get what they want to say, but it's not properly defined. "One example is the recent studies about stimulus-evoked..." The sentence explains that there are examples, but provides no citations! Also "One" and "exampleS"

Fig. 1: - The method to fit the power law is not detailed in the methods (just a vague reference to a package). This is a problem because some methods like least squares don't do well on power laws, and particularly for neuroscience due to low sampling (Wilting & Priesemann, Nat com.). - The "olive" curve is not "olive". Olive is dark green, and the color is purple. The problem appears in the subsequent figure.

Fig. 2: - The number of neurons is very small (19). This is very odd, since the original dataset has a lot of neurons. Also, the authors seem to pick age 97 and 102, but do not explain why those two points have any relevance. - If you run a correlation you need to explain what is the correlation (pearson, spearman?). It also matters where the variables are normalized or not, and there is no control for shuffling. - The authors mention "low dimensional", but don't explain what method they use (looks t-SNE to me). - The authors use the word "signal" while in the text they refer to the "mean activity". Are those the same? - "We reproduced previous results showing that low-dimensional embeddings of mean population response vectors for different signals remain similar across sessions" The blue and green clusters that the authors report as being close across sessions are not close. Red-green-grey seem to remain closer, but even that is quite a stretch. - Correlation across matrices is strange. Since the authors did not clarify the actual formula or method, the correlation of 0.5 in Fig. 2E could be simply due to the fact that all the variables are pre-selected to be positive (or above threshold). This would also have an important effect on the angle (Fig. G). In fact, it would explain how comes that the correlation does not decrease with Delta T (which is what would be expected from drift. - Whenever the authors run a statistical analysis, it would help to run a shuffled control.

Self-organised criticality emerges through homeostatic plasticity. - The authors refer a lot to reference 35, but it's not clear what is the difference between their work and that one. - The text provides a general overview and refers to the methods for details. Since most of the results are based on that mode, I suggest putting it in the main text (although this is an opinion, not a dealbreaker). - Especially, mention which populations are we talking about, what are the numbers of neurons in each, and how are they connected.

• Fig. 4 has a lot of the same weaknesses as Fig. 2. In fact, the results on E are very similar, despite the fact that the matrices in D are clearly not the same.

Enhanced Neural representation through self-organised criticality The phase transition seems to be an observation over a computational model, but I don't see much analysis. It would be nice to have some order parameter, although the plots are convincing without it. The authors do spend time talking about co-spiking and silent periods though, but don't actually plot this. The only reference is to S4, which actually only seems to cover the super-critical state.

Fig 6: - It might be true that the accuracy peaks at the critical point, but it's really hard to call it significant. The authors should run multiple models and assess significance. - I don't entirely see the point of C. What does it mean for the model? And although I assume it is on the same experimental data, the authors do not mention it.

Fig. 7: - Plot is squeezed, and has low resolution. - Since the authors didn't clarify whether they have II connections or not (some models use them, some don't), or whether their plasticity applies to inhibitory neurons, it is very hard to assess what are the differences between A and B.

References: There are a fair amount of works that studied computational models for criticality. I am particularly thinking of the works of Bruno del Papa "Criticality meets learning: Criticality signatures in a self-organizing recurrent neural network". Experimentally, there are works showing that the so-called spontaneous activity is actually very reliable (if you record enough neurons). Nghia et al. "Nguyen, Nghia D., et al. "Cortical reactivations predict future sensory responses." Nature 625.7993 (2024): 110-118."

An important point missing in this work is that it assumes that spontaneous activity is somehow intrinsically generated. This is not necessarily true of cortical areas (where it could easily come from hippocampus).

Reviewer #2 (Public review):

This work attempts to reconcile the concepts of critical neural dynamics with short-term reliable responses and long-term drifting responses. This is an important question, because critical dynamics are typically associated with unpredictable population responses to perturbations. Instead, this paper demonstrates that recordings from the mouse visual cortex include typical avalanche statistics in their spontaneous state as well as clustered within-session responses to natural movies. The authors find that a spiking neural network with homeostatic plasticity on inhibitory coupling captures the correlation-based metrics observed in experiments and that this network self-organizes into a critical state.

Strengths:

The structure of the manuscript is clear, and the line of argumentation is easy to follow. The question raised is valid, and the model employed to answer it is adequate. While I am unsure if representation should be equated with reliable responses, I find the framework of reliable responses well-suited to compare experimental and numerical data.

Weaknesses:

• The claim that the presented model "self-organizes to the critical spontaneous state" is incompatible with Fig. 6 showing that the inhibitory timescale is a control parameter of the transition from subcritical to supercritical avalanche statistics.

• The notion of "drift" implies to me a gradual change on long timescales. This is demonstrated in Ref. [47] for a model including two different types of plasticity. Also, such a drift over time was observed in Ref. [11] Fig.3C. In the present work, we can see from Fig. 2E that the correlation drops immediately to a plateau. Instead, the model actually shows some decay of correlations, expected from the ongoing plasticity. This challenges the claim that the "model successfully reproduce[s] both representational drift and [...]". Instead, the model of [47] does reproduce representation drift.

• The claim that "spontaneous self-organized criticality serves as [...] functional mechanism for maintaining reliable information representation under continuously changing networks" is not justified by the above-raised points.

• From the methods, I understand that the dimensionality reduction in Fig.2C and Fig.4C is a result of independent t-SNE. Since t-SNE to my knowledge starts with a random projection of data to then optimize the embedding, the resulting orientation of independent runs cannot be compared such that statements like "rotation of low-dimensional representations as in Fig. 2C, where nodes (centers of the same-color clusters) change their positions across sessions (top panel and bottom panel), but their relative positions remain stable" are not possible.

Reviewer #3 (Public review):

Summary:

This study uses computational modeling of a spiking network of E-I with homeostatic inhibitory plasticity and aims to show that self-organized criticality that arises from the homeostatic mechanism can result in representational drift as well as reliable stimulus representation, because the geometric representation of stimuli remains restricted.

Strengths:

This paper provides a framework to link critical spontaneous state, homeostatic inhibitory plasticity, representational drift, and stimulus population response reliability

Weaknesses:

The study does not show a causal (or necessary/ sufficient) relationship between criticality at the spontaneous state, representational drift, and reliable stimulus presentation. The study only reports an observation that these features could co-exist. However, it does not show how the criticality of the spontaneous state could restrict the manifold for stimulus response.