Behavioral timescale synaptic plasticity (BTSP) is a form of single-shot learning observed in hippocampal place cells in mice (Bittner et al., 2015, 2017). This finding is both interesting and inspiring for computational neuroscience for several reasons. In the first place, synaptic plasticity is the basis for learning, and any neuronal theory on the organizing principles of the brain must consider synaptic plasticity. In addition, previous findings on plasticity were usually based on artificial stimulation protocols in slices (LTP, LTD, STDP), often requiring unphysiological stimulation intensities, careful selection of stimulation protocols and many repetitions, while BTSP is a single-shot learning paradigm. Here are a some quite recent interesting theory papers on BTSP, complementing a previous blog post about experimental studies on BTSP:

Nonlinear slow-timescale mechanisms in synaptic plasticity

This review by O’Donnell (2023) nicely covers both experimental results and modeling work about the potential mechanisms of synaptic plasticity on a slow timescale. O’Donnell attempts to reconcile seeming discrepancies between experimental results on LTP from a theory perspective, and he stresses the importance of non-linearities to induce plasticity specifically during bursts but not regular spiking. A quick and interesting read!

One aspect I found particularly interesting is when he points out the relevance of long timescales and non-linear processes for classic slice LTP/LTD or STDP protocols (pages 2-3 of the PDF). Typically, one forgets about the long timescale and only focuses e.g. on the millisecond timing difference for the STDP protocol.

How BTSP affects subthreshold representations of place

Milstein et al. (2021) perform some really nice experiments and analyses to show that BTSP does not only induce new place fields but can alter existing synaptic weights in both ways, potentiation and depression.

If you read the abstract of the Milstein et al. study and have no idea what this is about: worry not, the rest of the paper is much easier to digest and actually super interesting. After reading the abstract myself, I was rather confused. One of the problems is that the “bidirectional plasticity” mentioned in the title can be misread as plasticity among bidirectionally connected cell pairs, while here the “bidirectional” only refers to “depression” (down-direction) and “potentiation” (up-direction) of synpases.

Two things I like in particular about this paper: Fig. 3L, which summarizes the main finding in a very nice visualization. And the experiments reported in Fig. 4A-K, where the authors try to disentangle the contributions of cellular depolarizations vs. initial synaptic weights for plasticity – a beautiful perturbation experiment.

P.S. This is not exactly a very “recent” paper, but I found it useful as context to better understand the subsequent paper by Moldwin et al. (2023).

A calcium level-based, general framework for plasticity

Moldwin et al. (2023) describe a framework for synaptic plasticity focused on calcium levels. It is based on the well-established finding that (a) synaptic depression occurs for low calcium levels and (b) synaptic potentiation occurs for high calcium levels, while (c) synapses remain approximately stable for the lowest (baseline) calcium levels. With their modeling work, the authors extend on existing models and make them more flexible and more easily parameter-adjustable.

In general, I think it is interesting to generate such a more general and flexible framework. At some points, however, I had the feeling that adding more model features to fit all use cases does not necessarily make the model better (more powerful yes, but not always more useful). For example, the authors write that some experimental edge cases were “incorporated into [their model] by adding additional thresholds into the step function”. Adding parameters to account for special cases does not seem the most elegant way if this additional complexity does not make any further predictions about model behavior. Despite these concerns, however, I believe that their general model based on calcium as a single actor is interesting, and their attempt to apply the model to diverse plasticity protocols is useful. It also helps to understand which assumptions are necessary to make the calcium-based plasticity model work for a particular experiment.

In the title, the authors claim that this framework could also be used to describe BTSP. Their take on BTSP is similar to the one described by Milstein et al. (2021) (see above).  Moldwin et al. (2023) assume that there is a summation of synaptic calcium induced by pre-synaptic stimulation and calcium induced by the plateau potential. The summed calcium signal would then induce LTP or LTD at this synapse. The main difference compared to Milstein et al. (2021) is therefore the idea that both slow signals, the eligibility trace and the instructive signal, are encoded in the local calcium concentration; apart from that, the mathematical shape of their model seems to be pretty similar compared to Milstein et al. (2021) as far as I can judge.

The main problem of this application to BTSP is the decay time of calcium at synaptic locations. The authors make the following assumption: “The synaptic calcium decayed with a time constant of ~2 seconds”. This seems really unrealistic – calcium levels after synaptic activation typically decay faster by an order of magnitude. Most experimentalists would assume that this long timescale is taken care of by a slow turnover of more stable mediators like CaMKII (for potentiation) or Calcineurin (for depression); see also the review by O’Donnell (2023). Overall, I have the impression that this framework is a nice and generalizable model for LTP, LTD and STPD, but maybe less so for BTSP. It is unclear to me why they advertise their model to cover BTSP so prominently in the title – this aspect may not represent the best part of their work.

A review on normative models on synaptic plasticity

In their review paper, Bredenberg and Savin (2023) discuss the criteria that should be used to judge normative models on synaptic plasticity. What do they mean with “normative models”? The authors nicely describe a very clear distinction between phenomenological and mechanistic models on the one side, which use fitting or equations to describe experimental observations; and normative models on the other side, which try to understand the importance of the modeled process for the organism. These normative models therefore ask: how should synaptic plasticity be like in order to serve the organism well? That’s an interesting perspective, not driven bottom-up by experimental observations but top-down from “desiderata”, as the title of this review indicates.

The review is interesting and establishes – not surprisingly given its focus on the practical usefulness of plasticity models – many connections to deep learning-based approaches. I was, however, slightly disappointed by the apparent lack of discussion of behavioral timescale synaptic plasticity (BTSP). For example, they describe how to choose an objective function for synaptic plasticity to improve performance and how to linearize it for small weight changes, which I found a very interesting point of view. It would be instructive to investigate BTSP and single-shot learning in such a model. In particular, it seems to me that BTSP – due to its single-shot learning behavior – cannot be a learning rule based on small and linear weight changes. It would therefore be interesting to describe BTSP in a normative model and, for example, to investigate how a loss function can be used to derive or describe BTSP.

.

Did I miss an interesting recent paper on this topic? Let me know!

.

.

References

Bittner, K.C., Grienberger, C., Vaidya, S.P., Milstein, A.D., Macklin, J.J., Suh, J., Tonegawa, S., Magee, J.C., 2015. Conjunctive input processing drives feature selectivity in hippocampal CA1 neurons. Nat. Neurosci. 18, 1133–1142. https://doi.org/10.1038/nn.4062

Bittner, K.C., Milstein, A.D., Grienberger, C., Romani, S., Magee, J.C., 2017. Behavioral time scale synaptic plasticity underlies CA1 place fields. Science 357, 1033–1036. https://doi.org/10.1126/science.aan3846

Bredenberg, C., Savin, C., 2023. Desiderata for normative models of synaptic plasticity. https://doi.org/10.48550/arXiv.2308.04988

Milstein, A.D., Li, Y., Bittner, K.C., Grienberger, C., Soltesz, I., Magee, J.C., Romani, S., 2021. Bidirectional synaptic plasticity rapidly modifies hippocampal representations. eLife 10, e73046. https://doi.org/10.7554/eLife.73046

Moldwin, T., Azran, L.S., Segev, I., 2023. A Generalized Framework for Calcium-Based Plasticity Describes Weight-Dependent Synaptic Changes in Behavioral Time Scale Plasticity. https://doi.org/10.1101/2023.07.13.548837

O’Donnell, C., 2023. Nonlinear slow-timescale mechanisms in synaptic plasticity. Curr. Opin. Neurobiol. 82, 102778. https://doi.org/10.1016/j.conb.2023.102778 (preprint)