In recent years, there has been a surge of interest in how dendrites and their complex geometry allow for integration of spatiotemporal patterns of input and transformation into neuronal response. [1, 2] Although sizeable amounts of synaptic input arrive at distinct branches, the dendritic tree must convert this composite signal into meaningful information, which will be then transferred to the soma and potentially evoke spiking activity. [3] While there has been considerable development in dendritic integration modeling in the last two decades, single-compartment neurons are still a hallmark of computational neuroscience and machine learning [2]; this level of abstraction, however, ignores components of dendritic integration that may be essential to properly represent neuronal dynamics, hence limiting the performance of the studied systems.
Some of these shortcomings were recently overcome by a two-compartment model that introduced a self-supervision rule within a single neuron to minimize information loss between dendritic synaptic input and somatic output spiking activity. [4] Networks composed of this neuron model could perform a variety of unsupervised temporal feature learning tasks such as chunking and blind source separation, usually performed by specialized networks with different learning rules. We wish to generalize this learning principle and develop a new framework in which dendritic trees have two or more compartments with hierarchical, linear-nonlinear integrations. [3] Here we investigate how can self-supervision be defined in this system and examine its accuracy when presented to the previously introduced temporal feature learning tasks. We expect that, by distributing the synaptic input into different compartments of the same neuron, our model can use delayed integration to differentiate similar temporal patterns that were previously indistinguishable.
Acknowledgments
This work was partly supported by KAKENHI (nos. 18H05213 and 19H04994). We would like to thank the Ministry of Education, Culture, Sports, Science and Technology (MEXT) of Japan and Okinawa Institute of Science and Technology Graduate University (OIST) for supporting M. M. C.
References
1. Stuart G J, Spruston N. Dendritic integration: 60 years of progress. NatNeurosice. 2015, 18, 1713–1721. 2. Häusser M, Mel B. Dendrites: bug or feature? Current Opinion in Neurobiology. 2003, 13(3), 372-383. 3. Ujfalussy B B, Makara J K, Lengyel M, Branco T. Global and Multiplexed Dendritic Computations under In Vivo-like Conditions. Neuron. 2018, 100(3), 579 - 592.e5 4. Asabuki T, Fukai T. Somatodendritic consistency check for temporal feature segmentation. Nat Commun. 2020, 11, 1554.
