@ChrisWLynn

**Christopher W. Lynn** @ChrisWLynn [2025-04-16](https://x.com/ChrisWLynn/status/1912502656108544454) In the mouse hippocampus and visual cortex, we find that direct dependencies capture 90% of a neuron's activity. This leaves only 10% for interactions between inputs and inherent noise [image] --- **Christopher W. Lynn** @ChrisWLynn [2025-04-16](https://x.com/ChrisWLynn/status/1912502657912107077) This means that real neurons are closely approximated \*quantitatively\* by the first artificial neuron proposed by McCulloch and Pitts in 1943 [image] --- **Christopher W. Lynn** @ChrisWLynn [2025-04-16](https://x.com/ChrisWLynn/status/1912502659552076064) With only a small number of direct input-output dependencies (no interactions between inputs) we are able to predict complex higher-order dependencies [image] --- **Christopher W. Lynn** @ChrisWLynn [2025-04-16](https://x.com/ChrisWLynn/status/1912502662035104031) Moreover, the inferred connection weights are 1. sparse, 2. heavy-tailed, 3. balanced, and 4. directed -- all key features observed in synaptic wiring between neurons [image] --- **Christopher W. Lynn** @ChrisWLynn [2025-04-16](https://x.com/ChrisWLynn/status/1912502664383934718) For much more, see the preprint: "Simple low-dimensional computations explain variability in neuronal activity" https://arxiv.org/abs/2504.08637 --- **Toviah Moldwin** @TMoldwin [2025-04-17](https://x.com/TMoldwin/status/1912771825861271587) Nice work. I recommend you also take a look at @DavidBeniaguev 's work (single neurons as deep networks) as well as mine (perceptron learning in model cortical pyramidal cells). --- **Mihaly Hanics** @HanicsResearch [2025-04-17](https://x.com/HanicsResearch/status/1912665579447542005) Always questioned the true algorithmic practicality of some ideas in DL that were "too strongly taken" from neuroscience - the sum+activation is one. Regardless of the biological processes, I do not see a reason why some other function on a layer of neurons couldn't work better