A BIOLOGICAL NEURON is a cell specialised for signalling. Its parts: DENDRITES branching fibres receiving signals from other neurons. A single neuron may have thousands. SOMA the cell body, where incoming signals are integrated AXON a single long fibre carrying the output signal away AXON TERMINALS the branching end of the axon SYNAPSE the junction between an axon terminal and another neuron's dendrite. Signals cross it chemically, via neurotransmitters. HOW IT COMPUTES: · dendrites receive signals that are EXCITATORY (raising the cell's electrical potential) or INHIBITORY (lowering it) · the soma sums them · if the potential exceeds a THRESHOLD, the neuron FIRES — it sends an ACTION POTENTIAL down the axon · firing is ALL-OR-NOTHING: the spike has a fixed size, and information is carried by the RATE and TIMING of spikes, not by their amplitude · after firing there is a REFRACTORY PERIOD during which it cannot fire again LEARNING happens by changing SYNAPTIC STRENGTH — how much influence one neuron has on another. That is the biological fact the whole field is built on, and it is what Hebb's rule (a later topic) formalises. THE SCALE: about 10^11 neurons in a human brain each with roughly 10^4 synapses, giving on the order of 10^14–10^15 connections a neuron fires at up to roughly 100–1,000 times per second signal propagation is MILLISECONDS — about a million times slower than a silicon transistor's switching time THE INTERESTING CONSEQUENCE OF THOSE LAST TWO FIGURES: brains are built from very slow components and outperform computers on perception. THE ADVANTAGE MUST THEREFORE COME FROM MASSIVE PARALLELISM, not from speed — roughly 10^11 processors operating at once, against a handful of fast cores. That observation is the original motivation for connectionism.

AN ARTIFICIAL NEURON abstracts the biological one to a small amount of arithmetic: inputs x₁ … xₙ, each multiplied by a WEIGHT wᵢ the weighted sum, plus a BIAS b, forms the NET INPUT net = Σ wᵢxᵢ + b an ACTIVATION FUNCTION f produces the output y = f(net) x₁ ──w₁──┐ x₂ ──w₂──┤ ┌──────────┐ ⋮ ├───▶ │ Σ then f│ ───▶ y xₙ ──wₙ──┤ └──────────┘ bias b ┘ THE CORRESPONDENCE, term by term: dendrites → the inputs synaptic strength → the WEIGHTS soma's summation → the weighted sum firing threshold → the activation function and bias axon output → y synaptic change → weight update during learning WHAT WAS KEPT — and this is the honest list of what the analogy actually delivers: · MANY SIMPLE UNITS rather than one complex processor · learning by ADJUSTING CONNECTION STRENGTHS · a nonlinear response to a summed input · DISTRIBUTED REPRESENTATION — no single unit holds a concept; the knowledge is in the pattern of weights · GRACEFUL DEGRADATION: damage to a few units degrades performance rather than destroying it, because the representation is distributed · PARALLELISM in principle WHAT WAS DISCARDED, and each difference matters: BIOLOGICAL ARTIFICIAL ───────────────────────────────────────────────────────────── spiking — discrete events in a continuous real-valued continuous time output per forward pass asynchronous; each neuron synchronous; the whole layer fires on its own computes at once ~10^11 units, 10^14 synapses thousands to billions of parameters, but a different organisation ~1,000 distinct neuron types one or two unit types chemical neurotransmitters, a single scalar weight dozens of kinds learning by local rules only BACKPROPAGATION, which sends an error signal BACKWARD along connections — no known biological mechanism does this connections grow and are the architecture is fixed pruned physically before training energy: about 20 watts a large training run consumes many megawatt-hours learns from few examples needs many examples ───────────────────────────────────────────────────────────── THE MOST IMPORTANT ROW IS BACKPROPAGATION. It is the algorithm that makes artificial networks trainable, and it requires each connection to know the derivative of a global error with respect to its own weight — information that would have to travel backward along the axon. NO BIOLOGICAL MECHANISM FOR THIS IS KNOWN, so the central technique of the field is explicitly non-biological. Anyone claiming these systems learn "the way the brain does" is wrong on the most consequential detail. THE ENERGY COMPARISON IS ALSO WORTH KEEPING: a brain does what it does on roughly the power of a dim light bulb. That gap is one of the clearest indications that the analogy is loose rather than deep.

THE PROPERTIES THAT FOLLOW FROM THE ARCHITECTURE, and why they matter: 1. LEARNING FROM EXAMPLES No rule needs to be stated. This is the answer to the knowledge-acquisition bottleneck of ACtE0904 — the knowledge is extracted from data rather than from an expert. 2. GENERALISATION A trained network responds sensibly to inputs it has never seen, provided they resemble the training data — which is the "future resembles the past" assumption from the machine learning topic. 3. TOLERANCE OF NOISE AND INCOMPLETE INPUT Because the response depends on a weighted sum over many inputs, corrupting a few has limited effect. This is exactly where symbolic rule systems are brittle. 4. DISTRIBUTED REPRESENTATION and GRACEFUL DEGRADATION Knowledge is not localised, so there is no single point of failure — and equally no single place to look to find out what the network knows. 5. NONLINEAR FUNCTION APPROXIMATION With enough hidden units, a network can approximate any continuous function to arbitrary accuracy — the UNIVERSAL APPROXIMATION THEOREM, covered in the multilayer perceptron topic. What the theorem does NOT say is that the weights can be FOUND, which is the practical difficulty. 6. PARALLEL IMPLEMENTATION The arithmetic is matrix multiplication, which is why GPUs transformed the field — the parallelism was in the model all along and only became exploitable when the hardware arrived. AND THE DISADVANTAGES, stated as plainly: · OPAQUE. The knowledge is a matrix of numbers, so the network cannot explain a decision — the property the AI applications topic identified as a regulatory requirement and the reason rule-based systems persist. · DATA-HUNGRY, and it inherits whatever bias the data holds · NO GUARANTEES: training may converge to a poor solution, and there is no proof of correctness · MANY DESIGN CHOICES — architecture, activation, learning rate — each requiring experiment · COMPUTATIONALLY EXPENSIVE to train · CONFIDENTLY WRONG on inputs outside the training distribution, including adversarial examples WHERE THIS SITS IN THE SUBJECT — the symbolic/connectionist contrast from the AI concepts topic, now concrete: SYMBOLIC AI explicit knowledge, interpretable, brittle at the edges, needs no data NEURAL NETWORKS learned knowledge, opaque, robust to noise, needs data THEY FAIL IN OPPOSITE PLACES, which is precisely why neuro-symbolic work exists and why the field did not simply abandon one for the other. A NOTE ON WHY THE BIOLOGICAL INSPIRATION STILL MATTERS despite the differences: it supplied the ARCHITECTURAL IDEA — many simple units, learning by weight change, distributed representation — at a time when the alternative assumption was that intelligence required explicit symbol manipulation. THE INSPIRATION WAS PRODUCTIVE EVEN THOUGH THE IMPLEMENTATION DIVERGED, which is the usual relationship between biology and engineering.