Astral Codex Ten

ANNs

4 min read

ANNs

Article

ANNs is a recurring concept in the Astral Codex Ten archive, appearing 2 times across 2 issues between November 08, 2022 and September 12, 2025. The archive places it in contexts such as “artificial neural networks … do not have to deal with conduction delays”; “After ANNs succeeded at pretty much every task we threw at them since the early 2010s, and people started to focus on ANN-based models of the brain”. It most often appears alongside A Change of Heart, Abraham, Adams.

Metadata

  • Category: Concepts
  • Mention count: 2
  • Issue count: 2
  • First seen: November 08, 2022
  • Last seen: September 12, 2025

Appears In

Source Context

Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.

Highlights From The Comments On Brain Waves
November 08, 2022 · Original source
However, the logic here was flawed from the start, because artificial neural networks are a mathematical model simulated on a computer and the properties of the computational substrate of a simulation don't carry through into the simulation. And so ANNs indeed do not have to deal with conduction delays and synchronization issues the same way that biological neural networks do, but the reason does not have anything to do with the low-level details of silicon chips.
Your Review: The Synaptic Plasticity and Memory Hypothesis
September 12, 2025 · Original source
For most people even vaguely familiar with neuroscience and the brain, this claim rings as trivially true. This idea is so entrenched, in fact, that I'd wager most people don't even know it has a name. (I didn’t, anyway.) The most popular quantitative models of the brain, artificial neural networks (ANNs), assume that the strengths of the connections between neurons, or “weights”, completely determine how networks behave. These are the things that are assumed to change during learning, maybe through an algorithm like backpropagation, or maybe through something that looks more like Hebbian learning. If these models learn something, or store a memory, it has to be through changes in weights.
Inline links: artificial neural networks, backpropagation, Hebbian learning
The rough history of the SPM hypothesis, then, goes something like this. Around 1900, there was still much we didn’t know about the nervous system (or biology, for that matter; remember that molecular biology really only got started in the 1950s), but even early workers like Ramón y Cajal recognized that connections between neurons could offer a useful substrate for modulating circuit behavior, and hence for learning and memory. This lead to the “plastic change hypothesis”, which took a more modern form around 1950 with the work of Hebb et al. When LTP and LTD were discovered in the 1970s, and appeared to successfully account for a variety of longer-term forms of learning, the stock of synaptic weight changes rose, and the stock of other known forms of plasticity (e.g., plasticity in single-neuron properties, and in rates of neuron creation and death) fell. By 2000, when the SPM hypothesis paper was published, workers had good reason to believe LTP and LTD explained most learning and memory phenomena of interest. Other forms of synaptic plasticity were discovered after LTP and LTD, like homeostatic plasticity and spike-timing-dependent plasticity, but the SPM hypothesis is general enough to accommodate them. After ANNs succeeded at pretty much every task we threw at them since the early 2010s, and people started to focus on ANN-based models of the brain, the SPM hypothesis became a dogma.
Inline links: homeostatic plasticity, spike-timing-dependent plasticity