The video argues that 2022 was highlighted as the “year of neural networks,” but points out that the classic perceptron view of a neuron—as a simple linear summator followed by a threshold—greatly underestimates the computational power of real biological neurons. It reviews how a neuron’s anatomy (dendrites, soma, axon) and its ion‑channel machinery (voltage‑gated Na⁺, K⁺, Ca²⁺ channels, NMDA receptors) give dendrites active, nonlinear processing abilities. These dendritic mechanisms enable phenomena such as back‑propagating action potentials, NMDA spikes, and dendritic calcium spikes that can selectively respond to specific input strengths, allowing a single dendrite to compute an XOR operation—a task that, in artificial networks, requires multiple layers. The discussion then shows that detailed biophysical models of cortical pyramidal neurons need deep convolutional networks (5‑8 layers) to replicate their input‑output behavior, underscoring that single neurons possess computational complexity comparable to multilayer artificial neural networks. The take‑away is that individual neurons are far more sophisticated than linear summators, and insights from neuroscience can inform and improve artificial network design.
1. 2022 can be called the year of neural networks.
2. In 2022 a language model was released that can write a surprisingly decent essay.
3. In 2022 an AI art generator was released.
4. In 2022 a platform that turns any image into Anime was released.
5. These developments were made possible by advances in artificial neural networks.
6. Artificial neural networks are computing systems modeled as networks of interconnected nodes that learn to solve problems by recognizing patterns in training data.
7. Online searches for “artificial neural networks” often return statements that they work like the brain.
8. Biological neurons are actually much more powerful than previously thought.
9. Individual neurons in the brain function essentially like full‑blown neural networks themselves, with high information‑processing capabilities.
10. Early neural networks were inspired by descriptions of biological neurons that were considered accurate at the time.
11. The birth of machine learning can be traced to 1943, when Walter Pitts and Warren McCulloch introduced the perceptron.
12. The perceptron was designed to mimic an individual nerve cell, acting as a simple summator and comparator.
13. A perceptron receives input numbers, multiplies them by weights, sums the products, and compares the sum to a threshold; if the sum exceeds the threshold it outputs 1.
14. Interconnecting many perceptrons creates a neural network.
15. Training a neural network means adjusting the input weights so that the network maps inputs to correct outputs.
16. Over the years researchers have invented various activation functions, network architectures, and efficient weight‑changing algorithms.
17. Because the nodes in artificial networks are still called neurons, many people believe biological neurons work exactly like perceptrons.
18. A typical neuron consists of dendrites, a soma (cell body), and an axon.
19. Neurons are electrically excitable cells that generate brief electrical pulses (action potentials) that travel to other neurons.
20. Electric charge in neurons is carried by ions such as Na⁺, K⁺, Cl⁻, and Ca²⁺ inside and outside the cell.
21. The lipid membrane separates the cell from the extracellular space and is normally impermeable to ions.
22. Neurons contain special protein ion channels that can open and close, allowing specific ions to cross the membrane.
23. By regulating ion flow through these channels, neurons control their membrane voltage; Na⁺ inflow depolarizes the membrane, while K⁺ outflow hyperpolarizes it.
24. Voltage‑gated ion channels open or close depending on the membrane potential.
25. An action potential is the all‑or‑none electrical signal used for neuronal communication.
26. Action potentials are initiated at the axon initial segment where many Na⁺ channels open when the membrane voltage exceeds a threshold.
27. Na⁺ influx depolarizes the membrane further, recruiting additional Na⁺ channels in a positive‑feedback loop.
28. The wave of Na⁺ channel opening propagates along the axon and is transmitted to downstream neurons via synaptic transmission.
29. After depolarization, Na⁺ channels close and K⁺ channels open, allowing K⁺ to leave the cell and restore the resting membrane voltage.
30. This sequence generates an action potential and conveys one bit of information to downstream neurons.
31. The thresholding operation in a neuron is implemented by voltage‑gated channels, making the perceptron analogy partially correct.
32. The main limitation of the perceptron lies in its handling of inputs, not its output.
33. Historically dendrites were viewed as passive cables that merely convey electrical signals to the soma.
34. In the passive cable model, dendrites sum incoming synaptic signals; the weight of each input depends on receptor number and distance from the soma.
35. Combined with a somatic threshold, this passive‑dendrite view provides a biological basis for the perceptron model.
36. However, dendrites are not purely passive; they contain voltage‑gated ion channels that give them active information‑processing properties.
37. Dendrites possess voltage‑gated Na⁺ channels similar to those in the axon, allowing back‑propagating action potentials that can influence postsynaptic sites.
38. Back‑propagating dendritic spikes contribute to synaptic plasticity by adjusting input weights.
39. Fast Na⁺ channels in dendrites can generate small depolarizations that transiently amplify synaptic inputs.
40. NMDA receptors require both membrane depolarization and neurotransmitter binding to open, acting as coincidence detectors.
41. NMDA channels are non‑selective for cations, permitting Ca²⁺ and Na⁺ influx, which is important for synaptic plasticity.
42. NMDA‑mediated depolarization (an NMDA spike) involves Ca²⁺ influx and lasts on the order of hundreds of milliseconds.
43. NMDA spikes enable dendrites to perform non‑linear integration of incoming signals.
44. Dendrites can discriminate the temporal order of incoming action potentials; sequential activation in one direction yields a different response than the reverse direction.
45. This sensitivity to order and velocity allows single neurons to process temporal patterns and generate sequence‑selective output.
46. NMDA spikes have been shown to enhance stimulus selectivity in the visual cortex of awake animals, contributing to behaviorally relevant computations.
47. In 2020, researchers led by Matthew Larkum published “Dendritic action potentials and computation in human layer 2/3 cortical neurons.”
48. They recorded simultaneous somatic and dendritic activity and discovered a new electrical response initiated at dendrites by sufficiently strong excitatory input, termed a dendritic calcium action potential.
49. Dendritic calcium spikes result from Ca²⁺ influx and have a shorter timescale than NMDA spikes.
50. These calcium spikes are highly selective to a particular input strength; too weak or too strong stimulation fails to elicit a spike.
51. This input‑strength selectivity enables a dendritic branch to implement an exclusive‑OR (XOR) operation: activating either set of synapses A or B alone triggers a spike, but activating both together does not.
52. Thus a single dendritic branch can compute the XOR function, a computation previously thought to require multi‑layered networks.
53. The biophysical basis for this selectivity involves voltage‑gated calcium channels and special potassium channels that are sensitive to both voltage and calcium concentration.
54. Recognizing that biological neurons are complex computational devices challenges the oversimplified assumptions underlying traditional artificial neural networks.
55. A study titled “Single cortical neurons as deep artificial neural networks” asked whether a deep network could replicate the input‑output transformation of a single cortical neuron.
56. The researchers built a detailed biophysical model of a cortical neuron using a reconstructed morphology and differential equations for membrane voltage, ion‑channel dynamics, and ionic fluxes.
57. They trained deep convolutional neural networks with varying numbers of layers to predict the model’s somatic voltage given the same synaptic inputs.
58. The deep network required between five and eight hidden layers to accurately predict the output spikes of the detailed model.
59. Removing NMDA channels from the biophysical model reduced the needed network complexity to a single hidden layer.
60. This demonstrates that dendritic non‑linearities, especially NMDA channels, contribute substantially to the computational power of neurons.
61. Remarkably, the deep network trained on random synaptic inputs could generalize and accurately predict outputs for spatially clustered, synchronous input patterns it had never seen before.
62. Thus the network inferred the underlying biophysics without explicit specification of those mechanisms.
63. The findings indicate that single cortical neurons, with their dendritic integrative properties, are sophisticated computational units comparable to a multi‑layered convolutional network.
64. Modeling a single neuron with an eight‑layer deep network is about 2000 times faster than simulating the detailed biophysical model, which requires solving many partial differential equations.