Neural Networks and Alternative Models

LaidLaid out as a single historical line — perceptron, then MLP, then the rest — this section's architectures would look like one lineage of fixes. They aren't. Weighted networks, RAM networks, prototype-based networks, attractor networks, margin-based classifiers, and instance-based models are substantially different traditions that developed in parallel, each making a different bet about what's worth keeping from training data. This page collects them into several different views side by side, because any single taxonomy flattens distinctions a reader actually needs.

By What Gets Stored FoundationalKnowledge that endures for decades — core principles

ThisThis is the organising question the first page of this section opened with, and it sorts every architecture covered here into a small number of genuinely different categories.

Stored state	Models
Continuous weights, fitted by gradient descent	Perceptron, MLP, RNN
Explicit, addressed memory contents	RAM / weightless networks
Prototype or centre vectors	RBF networks, Self-Organising Maps
Connection weights encoding attractor states	Hopfield networks
A subset of the training examples themselves, chosen by margin	Support Vector Machines
Every training example, unmodified	Nearest-neighbour models

By How Information Moves FoundationalKnowledge that endures for decades — core principles

Flow	Models
Feedforward — input to output, no cycles during inference	Perceptron, MLP, RBF network
Recurrent — output fed back as input, step by step	RNN
Recurrent until settled — feedback run to a stable point, not a fixed number of steps	Hopfield network
Addressed lookup — no propagation at all, one table read	RAM network
Neighbourhood competition — one winner, graded update to nearby units	Self-Organising Map

By How Training Actually Works FoundationalKnowledge that endures for decades — core principles

Mechanism	Models
Gradient-based optimisation of a loss function	Perceptron, MLP, RNN
Direct memory writes or count updates	RAM network
Competitive, neighbourhood-weighted updates	Self-Organising Map
Centre selection, then a direct least-squares fit	RBF network
Margin optimisation over a constrained set of candidates	Support Vector Machine
No fitting at all — retain the data, decide at query time	Nearest-neighbour model

NoNo single one of these three views is more "correct" than the others — each answers a different question about the same set of models, and a real system occasionally combines ideas from more than one row at once (an RBF network, for instance, is feedforward in information flow but fits its output layer the way a linear regression does, not the way an MLP's hidden layers do). That's more truthful than a single evolutionary tree running from perceptrons to transformers, which is exactly the oversimplification this section was built to correct.three views, not one tidy family tree

A Working Comparison FoundationalKnowledge that endures for decades — core principles

Model	Persistent state	Typical learning process	Key inductive bias
Perceptron	Weights and bias	Error-driven updates	Linear separation
MLP	Layer weights and biases	Backpropagation	Composed nonlinear functions
RAM network	Addressed memory values	Direct writes or count updates	Recognition from stored tuples
RBF network	Centres, widths, output weights	Centre selection, then least-squares fit	Local similarity
Self-Organising Map	Prototype vectors on a grid	Competitive neighbourhood updates	Topological organisation
RNN	Weights defining state transitions	Backpropagation through time	Sequential dependency
Hopfield network	Recurrent connection weights	Direct pattern storage	Attractor recall
Support Vector Machine	Support vectors and coefficients	Margin-maximising optimisation	Wide separating margin
k-Nearest-Neighbour	Stored examples	None — defer to query time	Nearby examples share an output

RealReal systems routinely combine several of these ideas rather than picking exactly one row — a modern image pipeline might use a convolutional network's learned features as the input to a much simpler nearest-neighbour or SVM classifier, for instance — so this table describes the pure, textbook version of each idea, not a claim that production systems always keep them this separate.

Related Topics

• What Is a Learning Model? — where the "what gets stored" question that organises this whole page was first raised.
• Symbolic and Connectionist Reasoning: Two Ways to Build "Think" — a different, complementary cut across the same territory, one level up from architecture choice.

Graduate Attributes

Directly developed

• Critical Thinker: Readers compare multiple classification frameworks to evaluate how different criteria reveal distinct aspects of neural network architectures.
• Subject Specialist: Readers distinguish between parallel traditions of network design rather than accepting a single linear historical narrative.

Modelled on this page

• Reflective: The limitations of simplified evolutionary trees are explicitly acknowledged to correct common misconceptions about model development.
• Confident Communicator: Technical distinctions between architectures are explained through clear, comparative categories tailored to different analytical questions.

University of Reading Graduate Attributes framework.