Hebrew University PNAS Study Shows Human Cortical Neurons Perform Deep Neural Network-Level Computations

The PNAS paper led by Idan Segev and Mickey London combined biophysical dendritic models with artificial neural networks to quantify computational depth. They trained ANNs to replicate input-output mappings of recorded human and rodent cortical neurons, finding human cells required networks with substantially greater hidden layers and parameters. This approach directly measures how dendritic morphology and ion-channel kinetics expand each cell's effective nonlinearity beyond the point-neuron simplification common in prior theory.

Prior coverage emphasized the raw result without noting that the method also supplies a falsifiable metric linking tree geometry to function, something earlier morphological studies lacked. The finding connects to 2016-2022 work showing human layer 2/3 pyramids possess slower Na+ kinetics and larger dendritic compartments than rodent homologs, yet stops short of testing whether these traits scale to network-level advantages in vivo. It reframes the long-running debate on human intelligence away from sheer neuron count toward per-cell algorithmic power.

The main limitation is reliance on ex-vivo slices and simulated inputs rather than behaving humans; causal impact on cognition therefore remains inferential. Longitudinal paired recordings in non-human primates combined with optogenetic perturbation of dendritic compartments would strengthen the claim that single-cell complexity causally supports higher cognition.

Future work should test whether transplanting human-like dendritic parameters into cortical organoids or neuromorphic hardware measurably improves few-shot learning or energy efficiency, providing a concrete bridge from cellular biophysics to both neuroscience and AI engineering.