Lab's latest out in @TrendsCognSci: "Discovering the Computational Relevance of Brain Network Organization". Unifies bio. & artificial neural nets via activity flow. Network coding models simulate how cognition emerges via connectivity. doi.org/10.1016/j.tics… (1/7) 
We realized even if all neurons' functions were mapped we still wouldn't know how the brain worked. Same goes for mapping all connections (connectome). We propose combining functional & connectivity maps, using predictive network models as "glue": network coding models. (2/7)
Remarkably, this "glue" involves the same algorithm as neural net simulations, such that insights into brain mechanisms can go along with unifying empirical & theoretical neuroscience. This activity flow algorithm is called the activation & propagation rules in neural nets. (3/7)
Many reasons we thought this shouldn't work, yet we proved ourselves wrong in two key studies (rdcu.be/kOJq & rdcu.be/wQ1M). Simulations showed activity flow prediction only works if connectivity determines functionality, yet it works empirically. (4/7)
The basis for network coding models was made even clearer in Ito et al. (2017) (rdcu.be/wQ1M), where we found we could use estimated activity flow over empirical connectivity patterns to predict localized neural activity patterns coding for cognitive information. (5/7)
Looking forward, we see advances in deep neural nets providing hypotheses that network coding models can test empirically. There are already good examples of this in vision neuroscience, which have utilized deep neural network models to predict function in visual cortex. (6/7)
We expect incorporation of empirical connectivity to advance this work via additional model constraints. This will make the models more like the human brain, likely advancing neural nets & understanding of neurocognitive functionality. Open access: arxiv.org/pdf/1907.03612… (7/7)
