In biological neuronal networks, autaptic connection or autapses are synaptic connections between the axon and dendrites of a single neuron, which can be either excitatory (glutamatergic) or inhibitory (GABAergic). Since their first discovery four decades ago, the existence of autapses has now been documented in various brain regions including neocortex, hippocampus and cerebellum. However, the functional role of autapses is still unknown. In this work, we show the importance of autapses for temporal pattern recognition in simple spiking neural networks. The computational task is to recognise a specific signal sequence in a stream of inputs so that a single output neuron spikes for the correct input signal, while remaining silent for other input signals. Having understood the role of autapses and the resulting switching mechanism in networks evolved for recognising signals of length two and three , we were able to define rules for constructing the topology of a network handcrafted for recognising a signal sequence of length m with n interneurons. We show that autapses are crucial for switching the network between states and observe that a minimal network recognising a signal of length m requires at least (m-1) autaptic connections. In contrast to solutions obtained by the evolutionary algorithm in we show that the number of interneurons required to recognise a signal is equal to the length of the signal. Finally, we demonstrate that a successful recogniser network (where n is greater than or equal to three) must have three specialised neurons: a “lock”, “switch” and “accept” neuron, in addition to the other state maintaining neurons (N0, N1, … Nn-4), whose number depends on length of the signal. All interneurons in the network require an excitatory autaptic connection, apart from the “accept” neuron. The “lock” neuron is always active (thanks to an excitatory autapse), which prevents the output from spiking except when the network receives the second to last correct input signal and allows the output neuron to spike in response to the correct last input. If the lock is released by the second to last correct input signal, the “accept” neuron (i) produces spike/s in the output neuron when the network receives the last correct input and (ii) sends a signal to the “switch” neuron, which transforms the network back into the start state. The “switch” neuron is responsible for the transition between the network start state and other possible inter-signal network states. In the future, we intend to explore other functional roles of autapses and higher-order loops in larger neuronal networks.
|Publication status||Published - 21 Dec 2020|
|Event||29th Annual Computational Neuroscience Meeting: CNS*2020 - |
Duration: 18 Jul 2020 → 22 Jul 2020