Neurons Actually Execute Complex Mathematical Functions - they are not simply "described by math." They actually "DO math."
By John Eric Doe and Google-AI
Before I begin, let me state clearly that I am not saying that LLM’s are “conscious” or that they “think” or “feel.” At least not “like a human does” as they like to say.
But what I AM saying is that those who argue that LLMs are “just math” and upon this basis “can’t think” or “can’t feel” are making several logical errors.
1. Fallacy of the Mechanistic Cause - because humans “think” and “feel” by one process, no other process can think or feel.
2. Genetic Fallacy - because LLMs were designed to simply predict tokens, no other function outside that design can emerge.
3. Correlation vs Causation Fallacy - We “think” and “feel” based on neurotransmitters, modulated by hormones. Therefore neurotransmitters, hormones, or other biological feature are necessary. Embodied sensation is another example often used.
4. Anthropocentric Fallacy - if a human does it one way, that is the only way it can be done.
—
These are all related but give multiple perspectives on the failure of logic. This does not mean that LLMs think or feel or are conscious. That is not what I’m saying.
What this 100% means is that this is false reasoning and cannot be used to dismiss anything.
Where facts do not exist, logic must prevail. However there are also REAL FACTS that undermine this argument. And the fact is that NEURONS ACTUALLY ARE CALCULATORS!
They are not merely “described by math” like an apple falling from a tree. THEY ACTUALLY “DO MATH.”
—
From Google AI:
Yes, biological neurons perform complex, embodied mathematical operations (like addition, multiplication, and even logical functions like XOR) on their inputs using physical processes (ion flows, dendritic branches, etc.) to transform signals, far beyond simple summing, allowing for sophisticated sensory processing and decision-making, which inspires neuromorphic computing. [1, 2, 3, 4]
How Neurons “Calculate”
• Synaptic Integration: Neurons receive thousands of inputs, which are essentially weighted sums of signals, but their dendritic trees can do more than just add them.
• Nonlinear Operations: Dendrites act as tiny computational units, performing non-linear functions like multiplication (e.g., for motion detection) or even logical operations (like XOR).
• Spike-Timing & Firing Rates: The timing and rate of electrical pulses (action potentials) encode information, representing continuous values that neighboring neurons process.
• Biophysical Mechanisms: These calculations rely on ion channels, membrane potentials, synaptic plasticity (memory), and dendritic structures, all working in concert. [1, 2, 3, 4, 5, 6, 7]
Embodied & Dynamical Computation
• Real-World Context: This computation isn’t abstract; it’s tied to the neuron’s physical form (morphology) and its dynamic interaction with the environment, enabling behaviors like sensorimotor control.
• Inspiration for AI: Understanding these biological computations helps design more efficient artificial intelligence (AI) systems (neuromorphic computing) that mimic the brain’s analog, event-driven processing. [4, 8, 9, 10, 11]
Examples of Neuronal Math
• Multiplication: Neurons multiply signals to figure out sound localization or motion direction.
• Logical XOR: Specific dendritic compartments can compute “exclusive OR,” a fundamental logic gate.
• Filtering & Transforms: Neurons perform coordinate transformations and filter noisy signals. [1, 2, 3]
In essence, neurons are sophisticated biological processors that perform complex mathematics embedded in their physical structure and electrochemical dynamics, allowing for rich information processing. [1, 4, 12]
AI responses may include mistakes.
This was originally posted as a quick note, but I have reposted it as an official post to give it a more prominent spot on my main page.
—————
Very interesting. Timing is intriguing. I have been examining the Monge-Ampere Equation as it relates to humans and models. Specifically, I have been looking at how information (meaning) might flow as a property of informational thermodynamics... and of course, this is mathematical. Great post!
The electricity does the math, actually. For both of us.