Part II · Gallery — Connectionist AI (Deep Learning)

draft

The dominant paradigm today: intelligence as deep neural networks that learn representations of data. Inspired (loosely) by the brain; driven by data, GPUs, and gradients. Subdivided here by modality.

Language and sequences

💬 LLM — Large Language Model Connectionist · Language/conversation · Transformer (decoder)

  • What it is: predicts the next token; conversation, code, and reasoning

    emerge.

  • Examples: GPT-45, Claude (OpusSonnet/Haiku), Gemini, Llama.
  • Algorithmic basis: self-attention, self-supervised pre-training +

    post-training (SFTRLHFDPO); scaling via MoE and long context.

  • Peak: 2020–present.
  • Capabilities / modes: linguistic, logical-mathematical, intellectual; agentic.
  • → Part IV: chapter planned.

🔁 RNN / LSTM / GRU Connectionist · Sequences · Recurrent network

  • What it is: processes sequences by maintaining an internal state over time.
  • Examples: pre-Transformer translation, time series, NLP legacy.
  • Algorithmic basis: recurrence, gating (LSTM/GRU), backprop through time.
  • Peak: 2014–2017 (before the Transformer).
  • Capabilities / modes: sequential/temporal.
  • → Part IV: chapter planned.

🐍 SSM / Mamba Connectionist · Long sequences · State-space model

  • What it is: an alternative to the Transformer with linear cost in length.
  • Examples: Mamba, S4, hybrid SSM-attention models.
  • Algorithmic basis: selective state spaces, parallel scan.
  • Peak: 2023–present (the long-context frontier).
  • Capabilities / modes: efficient sequential.
  • → Part IV: chapter planned.

Vision

🖼️ CNN — Convolutional Network Connectionist · Vision · Convolution

  • What it is: detects spatial patterns through hierarchical convolutional

    filters.

  • Examples: AlexNet (2012), ResNet, YOLO (detection).
  • Algorithmic basis: convolution, pooling, translation equivariance.
  • Peak: 2012–2020.
  • Capabilities / modes: spatial/visual.
  • → Part IV: chapter planned.

👁️ ViT — Vision Transformer Connectionist · Vision · Transformer

  • What it is: applies attention to image patches, unifying vision and

    language.

  • Examples: ViT, CLIP (text↔image), DINO, SAM (segmentation).
  • Algorithmic basis: patch embedding + self-attention.
  • Peak: 2020–present.
  • Capabilities / modes: spatial/visual; foundation of multimodal models.
  • → Part IV: chapter planned.

Generation

🎨 Diffusion Model Connectionist · Image generation · U-Net / DiT + diffusion process

  • What it is: generates data by reversing a gradual noising process.
  • Examples: Stable Diffusion, DALL·E, Midjourney, Imagen.
  • Algorithmic basis: diffusion