2nd/10_Wiki/Topics/AI_and_ML/Symbolic-AI vs Connectionism.md

---
id: wiki-2026-0508-symbolic-ai-vs-connectionism
title: Symbolic AI vs Connectionism
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [GOFAI vs Neural Networks, Logic vs Learning, Symbolic vs Subsymbolic]
duplicate_of: none
source_trust_level: A
confidence_score: 0.95
verification_status: applied
tags: [ai-history, symbolic-ai, connectionism, neuro-symbolic, philosophy-of-ai]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: pytorch+z3
---

# Symbolic AI vs Connectionism

## 매 한 줄
> **"매 symbolic 은 rules 의 manipulate, connectionist 는 weights 의 learn — 매 century-long debate"**. 매 1956 Dartmouth → 1980s expert system winter → 2012 AlexNet → 2022 ChatGPT 의 connectionist victory. 매 2026 의 답: 매 winner 없음, 매 hybrid (neuro-symbolic) 의 survive.

## 매 핵심

### 매 historical timeline
- **1956 Dartmouth**: McCarthy, Minsky, Newell, Simon → symbolic dominant.
- **1958 Perceptron**: Rosenblatt — connectionist 의 first.
- **1969 Minsky/Papert "Perceptrons"**: XOR critique — 매 first AI winter.
- **1980s Expert Systems boom + bust**: MYCIN, knowledge engineering bottleneck.
- **1986 Backprop (Rumelhart)**: connectionist revival.
- **2006 Deep Belief Net (Hinton)**: deep learning awakening.
- **2012 AlexNet**: ImageNet 의 connectionist domination 의 시작.
- **2017 Transformer**: attention-based 의 begin.
- **2022 ChatGPT**: scale 의 power 의 evidence.
- **2024 AlphaProof / AlphaGeometry**: hybrid 의 IMO-level.
- **2026 현재**: pure-symbolic ≈ niche, hybrid mainstream.

### 매 symbolic 진영 (GOFAI)
- **표현**: discrete tokens, logic, rules, KG.
- **추론**: deduction, search, unification.
- **장점**: interpretable, compositional, sample-efficient on structured.
- **단점**: brittle, perception fail, knowledge-engineering bottleneck.
- **대표**: SHRDLU, Cyc, Prolog, expert systems, Z3.

### 매 connectionist 진영
- **표현**: distributed embedding, weight tensor.
- **추론**: forward/backward pass, attention, gradient descent.
- **장점**: learns from raw data, perception, generalization.
- **단점**: black box, hallucination, sample-hungry, OOD fragile.
- **대표**: perceptron, CNN, LSTM, Transformer, LLM.

### 매 fundamental tensions
- **Compositionality**: symbolic 의 native, connectionist 의 emergent (debated).
- **Systematic generalization**: Marcus critique 의 핵심.
- **Sample efficiency**: symbolic ≪ connectionist data hunger.
- **Grounding**: connectionist 의 native, symbolic 의 needs perception layer.

### 매 modern synthesis (neuro-symbolic, 2026)
- **Pattern A (LLM-as-coder)**: LLM 가 Python/Lean code 의 generate, symbolic engine 가 execute.
- **Pattern B (NN as perception, symbolic as reasoner)**: AlphaGeometry — NN proposes constructions, DD+AR proves.
- **Pattern C (differentiable logic)**: Scallop, DeepProbLog — gradient through logic.
- **Pattern D (RAG with KG)**: GraphRAG, Microsoft 2024 — embedding + KG triples.

### 매 응용
1. **AlphaProof**: LLM (Gemini) + Lean 4 → IMO 2024 silver.
2. **AlphaGeometry**: NN constructions + symbolic deduction → IMO geometry gold.
3. **GraphRAG**: KG-augmented retrieval — connectionist embed + symbolic graph traversal.

## 💻 패턴

### 1. Pure symbolic (Prolog-style)
```prolog
parent(tom, bob).
parent(bob, alice).
ancestor(X, Y) :- parent(X, Y).
ancestor(X, Y) :- parent(X, Z), ancestor(Z, Y).

?- ancestor(tom, alice).  % true
```

### 2. Pure connectionist (Transformer)
```python
import torch.nn as nn
class Tiny(nn.Module):
    def __init__(self):
        super().__init__()
        self.emb = nn.Embedding(50000, 512)
        self.tr = nn.TransformerEncoderLayer(512, 8)
        self.head = nn.Linear(512, 50000)
    def forward(self, x):
        return self.head(self.tr(self.emb(x)))
```

### 3. Hybrid: LLM proposer + Z3 verifier
```python
import anthropic, z3

client = anthropic.Anthropic()

problem = "Find x, y in [0,100] s.t. 3x + 2y = 47, x*y is prime."

resp = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=500,
    messages=[{"role": "user", "content":
        f"Output Z3 Python code (no prose) for: {problem}"}]
).content[0].text

# Execute LLM-generated symbolic code
exec_globals = {"z3": z3}
exec(resp, exec_globals)  # symbolic solver gives ground truth
```

### 4. AlphaGeometry-style construct + verify
```python
def alphageometry_step(problem, llm, dd_engine):
    while not dd_engine.solved(problem):
        construction = llm.suggest_aux_construction(problem.state)
        problem.add(construction)
        dd_engine.deduce(problem)  # symbolic forward chain
    return problem.proof
```

### 5. GraphRAG (hybrid retrieval)
```python
def graph_rag(query, kg, vector_store):
    # connectionist: semantic match
    docs = vector_store.search(query, k=20)
    # symbolic: extract entities + walk KG
    entities = extract_entities(query)  # NER (NN) → symbol
    subgraph = kg.k_hop_neighbors(entities, k=2)
    # combine
    return llm.answer(query, context=docs + subgraph.to_text())
```

### 6. Differentiable logic (Scallop sketch)
```python
import scallopy
ctx = scallopy.ScallopContext()
ctx.add_relation("edge", (int, int))
ctx.add_rule("path(x, y) :- edge(x, y)")
ctx.add_rule("path(x, y) :- edge(x, z), path(z, y)")

# NN outputs probabilistic edges; loss flows back through reasoning
ctx.add_facts("edge", [(0, 1, 0.9), (1, 2, 0.7)])
ctx.run()
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Perception 의 dominant (vision, audio) | Connectionist |
| Logical guarantees 의 필요 | Symbolic verify layer |
| Mixed (proof, planning) | Neuro-symbolic hybrid |
| Tabular small | Tree (gradient boosting) |
| Knowledge-rich QA | Connectionist + KG RAG |
| Code/math | LLM proposer + interpreter/Lean/Z3 verifier |

**기본값**: 매 LLM (connectionist) + verifier (symbolic) hybrid 의 pragmatic default.

## 🔗 Graph
- 부모: [[Philosophy-of-AI]]
- 변형: [[Symbols]] · [[Neural-Symbolic-Integration|Neuro-Symbolic-AI]]
- 응용: [[GraphRAG]]

## 🤖 LLM 활용
**언제**: history explain, position survey, hybrid pattern design.
**언제 X**: 매 ground truth math/logic — verifier 의 필수.

## ❌ 안티패턴
- **Pure symbolic 의 modern build**: 매 brittleness — knowledge engineering bottleneck.
- **Pure connectionist 의 logical task**: 매 hallucination — verifier 의 add.
- **Hybrid 의 over-engineer**: 매 simple task 의 simple model 으로 충분.
- **"Connectionist won" claim**: 매 incomplete — IMO-level 의 hybrid 의 need.

## 🧪 검증 / 중복
- Verified (Marcus "The Next Decade" 2020, Bengio·Russell·Hinton statements 2024, AlphaProof Nature 2024, Hinton Turing lecture).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — symbolic vs connectionist (history + modern hybrid) |