Different networks, different teaching analogies
A convolutional net does not do what a transformer does. Each architecture is built for a
different kind of input and a different kind of output. To make that stick, each card pairs
the network with a brain-inspired metaphor
— a teaching aid, not a biology lesson.
Tap a card to flip between the network side
and the analogy side . Multimodal systems
stitch several of these together — vision here, language there — because no single network
covers every data type.
Neural network Convolutional Neural Network Stacked filters that hunt for local patterns — edges first, then shapes, then objects.
Tap to see the analogy side → Teaching analogy Visual Cortex (V1 → V4) Historical teaching analogy: early vision areas like V1 respond to edges; V4 adds color and form. CNNs borrowed this layered idea as inspiration — not a one-to-one biological map.
Tap to see the network side → Neural network Vision Transformer Treats an image as patches and relates them globally — whole-scene context, not just local filters.
Tap to see the analogy side → Teaching analogy Holistic recognition (IT analogy) Loose analogy only: inferotemporal cortex is often discussed alongside whole-scene and face recognition. ViTs relate patches globally — a similar job on paper, not the mechanism IT uses.
Tap to see the network side → Neural network Transformer Network Attention over tokens — the architecture behind large language models. Needs serious compute at scale.
Tap to see the analogy side → Teaching analogy Language & Reasoning Function-level analogy: transformers process token sequences and sit at the base of modern LLMs. The parallel is what the system does — language, inference, planning — not a named cortical region.
Tap to see the network side → Neural network Probabilistic Neural Network Outputs a distribution, not a single answer. Trained to estimate probability density under Bayesian principles.
Tap to see the analogy side → Teaching analogy Uncertainty & Prediction Pedagogical analogy: instead of one hard guess, the network keeps a spread of plausible outcomes — like weighing odds before you commit to an answer.
Tap to see the network side → Neural network Multimodal AI Model Different neural networks handle different data types — vision, language, audio — then meet in a shared space.
Tap to see the analogy side → Teaching analogy Integrated Perception Everyday analogy: sight, sound, and speech feed one experience. Multimodal models route each data type through its own encoder, then merge in shared space.
Tap to see the network side →
Analogies inspired by public lecture framing. None of these pairings claim to describe how the brain actually works.