Periodic Reporting for period 4 - SynapSeek (Learning the shape of synaptic plasticity rules for neuronal architectures and function through machine learning.)
Berichtszeitraum: 2023-08-01 bis 2024-05-31
How do we learn to dance, play an instrument, or a game as complex as chess or go? How do we make a memory? The common answer to these questions is “through synaptic plasticity”, through changing the synaptic connectivity of neural circuits so that representative brain activity can be reliably triggered. Such connectivity changes are governed by rules, i.e. synaptic mechanisms which monitor the activity of their environment and stereotypically strengthen or weaken synapses accordingly. The shape and mode of operation of these rules is still largely unknown: For the more than hundred different connection types in cortical circuits, only a handful of rules has been described at all. Similarly, testing observed rules in simulations of cortical function has only seen limited success. Our slow progress is due to the extraordinary difficulty of measuring and observing synapses without interference.
Here, we follow a new approach. We are utilizing the growing power of machine learning methods to deduce synaptic plasticity rules directly. Newly developed search algorithms and sheer computational power allow us to integrate published data and infer synaptic rules in silico. We (1) develop a new mathematical expression of synaptic plasticity rules, experimentally appropriate and flexible enough to be implemented in a Machine Learning framework, dubbed SYNAPSEEK. We then (2) apply SYNAPSEEK to deduce the rules for building various neural structures with increasing complexity. Finally (3), we incorporate additional, experimental constraints to SYNAPSEEK to develop synaptic rules that shape network function as much as its structure. Our work establishes, for the first time, canonical sets of synaptic plasticity rules, based on the circuit structure they must produce, and the function they are meant to support. SYNAPSEEK promises wide ranging applications, from a basic understanding of cortical development to better protocols for Deep Brain Stimulation.
Here, we follow a new approach. We are utilizing the growing power of machine learning methods to deduce synaptic plasticity rules directly. Newly developed search algorithms and sheer computational power allow us to integrate published data and infer synaptic rules in silico. We (1) develop a new mathematical expression of synaptic plasticity rules, experimentally appropriate and flexible enough to be implemented in a Machine Learning framework, dubbed SYNAPSEEK. We then (2) apply SYNAPSEEK to deduce the rules for building various neural structures with increasing complexity. Finally (3), we incorporate additional, experimental constraints to SYNAPSEEK to develop synaptic rules that shape network function as much as its structure. Our work establishes, for the first time, canonical sets of synaptic plasticity rules, based on the circuit structure they must produce, and the function they are meant to support. SYNAPSEEK promises wide ranging applications, from a basic understanding of cortical development to better protocols for Deep Brain Stimulation.
The project continues to proceed very much as envisioned and we continue to make good progress on the realization of all Aims. In particular, we have seen the continued success of our framework in analysing and modifying synaptic plasticity rules with ML methods (Aim 1). Further, we have begun to test and expand our techniques to infer plasticity rules in spiking networks as well, i.e. we started to fulfill the expectations of Aim 2 and 3, as hoped. Our work up to the end of period 4 has led to 15 poster presentations (NCCD’19, COSYNE ’20,21,22,23, 24, BCCN 24-25, Neuromatch ’20, CNS2023, NeurIPS20,23, IST retreat ’20, Oxford Neuroscience symposium ’19). Additionally, we have 5 manuscripts at NEURIPS 20,23 (which was accepted in Period 4, at the time of writing this report), eLife, Plos Comp.Neurosc. and others. We have started new work, on Aim 3, including experimental data to constrain our automated search methods. We have successfully hired two graduate students, and two postdocs. We will submit abstracts to BCCN and Cosyne 2025.
The project is developing as expected and we are happy to report that we have substantially expanded the boundaries of the state of the art, such that our current methods are the most cutting-edge approaches to finding new plasticity rules. We are very optimistic that most of the Aims as described in the original grant can be fulfilled.