AI visualisation

Veena's AI components separate semantic models from their rendered views. IDs and named axes remain stable while values and focus change.

Neural networks

const model = neuralNetwork([
  { id: "input", label: "Input", nodes: ["x₁", "x₂", "x₃", "x₄"] },
  { id: "hidden", label: "Hidden", nodes: ["h₁", "h₂", "h₃"] },
  { id: "output", label: "Output", nodes: ["ŷ₁", "ŷ₂"] },
]);

const network = NeuralNetwork(model);
this.add(network);

Networks support named layers, stable node/edge IDs, dense weights, explicit sparse or skip connections, activation snapshots, and signal flow over unique edges.

Biological neurons

BiologicalNeuron is a stylised teaching view with three independently animated phases: input pulses travel inward through dendrites, charge illuminates the soma, and a signal travels outward along the axon to its terminals.

const neuron = BiologicalNeuron({ dendrites: 9, myelinated: true });
this.add(neuron);

timeline.animate(neuron).duration(1.2).to(neuron.receive());
timeline.animate(neuron).duration(0.5).to(neuron.illuminate());
timeline.animate(neuron).duration(1.5).to(neuron.fire());

The component is deliberately a visual explanation rather than a physiological simulation. Pair it with NeuralNetwork to explain which ideas inspired artificial neurons while keeping their differences explicit.

Semantic tensors

tensorAxis and tensorSnapshot attach meaning to each dimension and element. Operations preserve or transform that meaning:

• tensorAdd, tensorSplit, tensorMerge, tensorReshape
• tensorProject2d, tensorTranspose2d, tensorMatmul
• tensorScale, tensorCausalMask, tensorSoftmax
• tensorSampleCategorical

TensorView renders a chosen 2D projection while fixed named axes select the remaining slice.

Attention and transformers

TokenRow, AttentionMatrix, and TransformerBlock provide reusable teaching views. attentionMatrices() computes Q/K products, scaling, masking, and softmax for a complete small example rather than displaying arbitrary numbers.

Context and generation

ContextWindow visualises token-budget composition. KvCache, NormalizationPanel, NextTokenBoard, and EntropyBar cover common LLM mechanics and sampling stories.

See the focused examples under examples/ai/, especially self-attention-lesson.ts, tensor-semantics.ts, neural-networks.ts, and context-window.ts.