{
  "graph": {
    "nodes": [
      {
        "id": "gradient_descent",
        "kind": "concept",
        "label": "Gradient Descent",
        "summary": "Iterative optimization that steps parameters downhill along the loss gradient."
      },
      {
        "id": "backpropagation",
        "kind": "concept",
        "label": "Backpropagation",
        "summary": "Reverse-mode autodiff that computes loss gradients through a network."
      },
      {
        "id": "chain_rule",
        "kind": "concept",
        "label": "Chain Rule",
        "summary": "Calculus rule for differentiating composed functions."
      },
      {
        "id": "neural_network",
        "kind": "concept",
        "label": "Neural Network",
        "summary": "Layered function approximator built from weighted connections."
      },
      {
        "id": "loss_function",
        "kind": "concept",
        "label": "Loss Function",
        "summary": "Scalar objective that quantifies prediction error."
      },
      {
        "id": "learning_rate",
        "kind": "concept",
        "label": "Learning Rate",
        "summary": "Step-size hyperparameter scaling each gradient update."
      },
      {
        "id": "overfitting",
        "kind": "concept",
        "label": "Overfitting",
        "summary": "When a model memorizes noise instead of generalizable signal."
      },
      {
        "id": "regularization",
        "kind": "concept",
        "label": "Regularization",
        "summary": "Techniques that penalize complexity to improve generalization."
      }
    ],
    "edges": [
      {
        "id": "e1",
        "source": "chain_rule",
        "target": "backpropagation",
        "type": "prerequisite_of",
        "label": "underlies"
      },
      { "id": "e2", "source": "backpropagation", "target": "neural_network", "type": "explains", "label": "trains" },
      {
        "id": "e3",
        "source": "gradient_descent",
        "target": "loss_function",
        "type": "related_to",
        "label": "minimizes"
      },
      {
        "id": "e4",
        "source": "loss_function",
        "target": "gradient_descent",
        "type": "prerequisite_of",
        "label": "defines objective for"
      },
      { "id": "e5", "source": "learning_rate", "target": "gradient_descent", "type": "related_to", "label": "tunes" },
      {
        "id": "e6",
        "source": "gradient_descent",
        "target": "backpropagation",
        "type": "related_to",
        "label": "consumes gradients from"
      },
      { "id": "e7", "source": "neural_network", "target": "overfitting", "type": "related_to", "label": "can suffer" },
      {
        "id": "e8",
        "source": "regularization",
        "target": "overfitting",
        "type": "contradicts",
        "label": "counteracts"
      },
      {
        "id": "e9",
        "source": "regularization",
        "target": "neural_network",
        "type": "related_to",
        "label": "constrains"
      },
      {
        "id": "e10",
        "source": "loss_function",
        "target": "neural_network",
        "type": "related_to",
        "label": "evaluates"
      },
      {
        "id": "e11",
        "source": "learning_rate",
        "target": "loss_function",
        "type": "related_to",
        "label": "affects descent on"
      },
      { "id": "e12", "source": "chain_rule", "target": "gradient_descent", "type": "related_to", "label": "supports" },
      {
        "id": "h1",
        "source": "learning_rate",
        "target": "overfitting",
        "type": "related_to",
        "label": "hypothesis: high rate may worsen",
        "properties": { "edge_role": "hypothesis" }
      }
    ],
    "claims": [
      {
        "id": "c1",
        "claim_type": "note",
        "target_id": "gradient_descent",
        "text": "The optimization workhorse behind most deep learning training.",
        "origin": "user_asserted",
        "support_state": "source_supported",
        "review_state": "accepted"
      },
      {
        "id": "c2",
        "claim_type": "note",
        "target_id": "backpropagation",
        "text": "Makes gradient computation efficient via reverse-mode automatic differentiation.",
        "origin": "user_asserted",
        "support_state": "unverified",
        "review_state": "pending"
      },
      {
        "id": "c3",
        "claim_type": "note",
        "target_id": "overfitting",
        "text": "Gemma inferred this is a relevant failure mode for the networks discussed here — a hypothesis awaiting your review.",
        "origin": "model_inferred",
        "support_state": "unverified",
        "review_state": "pending"
      },
      {
        "id": "n1",
        "claim_type": "note",
        "target_id": "regularization",
        "text": "Dropout and weight decay are the most common regularization techniques here.",
        "origin": "user_asserted",
        "support_state": "unverified",
        "review_state": "pending"
      },
      {
        "id": "hyp1",
        "claim_type": "hypothesis",
        "target_id": "learning_rate",
        "text": "Hypothesis: an overly large learning rate amplifies overfitting on small datasets.",
        "origin": "user_asserted",
        "support_state": "unverified",
        "review_state": "pending"
      },
      {
        "id": "hyp2",
        "claim_type": "hypothesis",
        "target_id": "h1",
        "text": "Hypothesis edge: learning rate and overfitting may be causally linked.",
        "origin": "model_inferred",
        "support_state": "unverified",
        "review_state": "pending"
      },
      {
        "id": "q1",
        "claim_type": "question",
        "target_id": "loss_function",
        "text": "Open question: which loss best balances calibration and accuracy for this task?",
        "origin": "user_asserted",
        "support_state": "unverified",
        "review_state": "pending"
      },
      {
        "id": "q2",
        "claim_type": "question",
        "target_id": "chain_rule",
        "text": "Open question: how to explain the chain rule intuition without calculus prerequisites?",
        "origin": "user_asserted",
        "support_state": "unverified",
        "review_state": "pending"
      }
    ],
    "metadata": {
      "schema_version": "0.1.0",
      "node_count": 8,
      "edge_count": 13
    }
  },
  "scene": {
    "scene_id": "annotated_graph_initial",
    "visible_node_ids": [
      "gradient_descent",
      "backpropagation",
      "chain_rule",
      "neural_network",
      "loss_function",
      "learning_rate",
      "overfitting"
    ],
    "visible_edge_ids": ["e1", "e2", "e3", "e4", "e5", "e6", "e7", "e10", "e11", "e12"],
    "fogged_node_ids": ["regularization"],
    "highlighted_ids": [],
    "selected_id": "",
    "callouts": [
      { "id": "co1", "target_id": "learning_rate", "text": "Open question attached: tuning vs. overfitting." }
    ],
    "node_positions": {
      "gradient_descent": [0.0, 0.0],
      "backpropagation": [200.0, -80.0],
      "chain_rule": [400.0, -40.0],
      "neural_network": [120.0, 180.0],
      "loss_function": [-180.0, 120.0],
      "learning_rate": [-120.0, -160.0],
      "overfitting": [360.0, 150.0]
    }
  }
}