Fig 1.
A) Illustration of the biophysically-detailed model with 639 compartments of a cortical layer V pyramidal cell model [29], which is the main object of study in this paper, color-coded by the compartment depth (consistent across all panels). B) Membrane voltages as calculated by a biophysically-detailed simulation of the multi-compartment model, used as the ground truth throughout this paper. C) Membrane voltages as predicted by our best-performing multi-task learning architecture, one time step at a time with a time resolution of 1 ms for the prediction. D) Comparison between the ground truth and predicted extracellular potentials as calculated at six points outside the neuron (orange dots in panel A).
Fig 2.
Swarm plots of the generalisation loss for each of the compartments (basal, oblique, apical) of the neuron model for each of the three models: Hard Parameter Sharing (MH), Multi-gate Mixture-of-experts (MMoE), and Multi-gate Mixture-of-Experts with Exclusivity (MMoEEx).
Note that basal dendrites of a neuron receive incoming signals from other neurons and convey them towards the cell body, while the apical dendrite extends from the cell body to integrate signals from distant regions, and oblique dendrites play a role in the integration of synaptic inputs at various angles away from the other dendrites. The vertical axis represents the numeric loss values while the horizontal axis, the different ANN models are indicated. The density of the data points in a specific region indicates how many compartments have a similar loss value for each model.
Fig 3.
A) Illustration of a biophysically-detailed model of a cortical layer V pyramidal neuron [29], color-coded by the compartment depth (consistent across all panels). B) Membrane voltages of a back-propagating action potential activated Ca2+ spike as calculated by a biophysically-detailed simulation of the multi-compartment model. Note the presence of the characterizing features such as the long depolarization of around the main bifurcation point of the apical dendrite (yellow) and the two somatic (blue) action potentials that follow. C) Membrane voltages of a back-propagating action potential activated Ca2+ as predicted by our best-performing multi-task learning architecture, one time step at a time with a time resolution of 1 ms for the prediction.
Fig 4.
Mean weights from the experts in both Multi-gate Mixture-of-experts (MMoE; top row) and Multi-gate Mixture-of-experts with Exclusivity (MMoEEx; bottom row) after training the models on neural data, projected onto the different compartments of the cortical layer V biophysically-detailed neuron model.
Each row consists of five subplots, representing different experts within the models. The color scale (normalised to the mean weights of the first expert) is indicated by the horizontal color bar located below each row.
Table 1.
Inference speed of all three MTL architectures—Hard parameter sharing (MH), Multi-gate Mixture-of-experts (MMoE), and Multi-gate Mixture-of-experts with Exclusivity (MMoEEx)—After training on neural data generated by a biophysically-detailed model of a cortical layer 5 pyramidal neuron model compared to the classical NEURON simulation.
Fig 5.
Schematic representation of the architectures of the hard parameter model (left), the MMoE model (middle), and the MMoEEx model (right).