Amparo Gilhuis, Shirin Dora, Cyriel Pennartz, Jorge Mejias The brain continuously processes sensory information from multiple modalities, giving rise to internal representations of the outside world. If and how the information from multiple modalities is being integrated has extensively been investigated over the past years, leading to more insight in multisensory integration (MSI) and its underlying mechanisms [1]. However, the different experimental paradigms used to investigate MSI involve different cognitive resources and situational demands. In this study, we investigated how different experimental paradigms of MSI reflect on behavior output and in their corresponding neural activity patterns. We did so by designing a recurrent neural network (RNN) with the biological plausible feature of differentiating between excitatory and inhibitory units [2]. For each of the three multisensory processing tasks considered [3, 4], an RNN was optimized to perform the tasks with similar performance as found in animals. Network models trained on different experimental paradigms showed significant distinct selectivity and connectivity patterns. Selectivity for both modality and choice was found in network models that were trained on the paradigm that involved higher cognitive resources. Network models trained on paradigms that involve more bottom-up processes mostly experienced choice selectivity. Increasing the level of network noise in network models that at first did not experience modality selectivity led to an increase in modality selectivity. We propose that a higher range of selectivity arises when a task is more demanding, either due to higher network noise (which makes the task harder for the animal) or a more difficult experimental paradigm. The higher range of selectivity is thought to improve the flexibility of the network model, which could be a necessity for the network models to achieve good performance, and the resulting neural heterogeneity could be used for more general information processing strategies [5, 6].
Acknowledgements
This work was partially supported by the European Union’s Horizon 2020 Framework Programme for Research and Innovation under the Specific Grant Agreement No. 785907.
References
1) Chandrasekaran C. Computational principles and models of multisensory integration. Curr. Op. Neurobiol. 2017, 43, 25-34.
2) Song HF, Yang GR, Wang X-J. Training excitatory-inhibitory recurrent neural networks for cognitive tasks: a simple and flexible framework. PloS Comput. Biol. 2016, 12, e1004792.
3) Raposo D, Kaufman MT, Churchland AK. A category-free neural population supports evolving demands during decision-making. Nat Neurosci. 2014, 17, 1784.
5) Mejias JF, Longtin A. Optimal heterogeneity for coding in spiking neural networks. Phys. Rev. Lett. 2012, 108, 228102.
6) Mejias JF, Longtin A. Differential effects of excitatory and inhibitory heterogeneity on the gain and asynchronous state of sparse cortical networks. Front. Comput. Neurosci. 2014, 8, 107.
