Physical Symbol Systems & Language of Thought

TAKEAWAYS:

• Physical Symbol Systems (PSS) are computational models where cognition arises from manipulating symbols via formal rules, inspired by human reasoning.
• Language of Thought (LOT) proposes that cognition operates using an internal symbolic language, bridging mental processes and natural language.
• Symbol grounding is the challenge of linking abstract symbols to real-world perceptions and actions.
• Advantages of PSS: Explainability, logical consistency, and ease of implementation in AI (e.g., rule-based systems).
• Limitations of PSS: Struggle with perception, embodiment, and dynamic real-world interactions (e.g., robotics).
• Neural networks (distributed processing) contrast with PSS by excelling in pattern recognition but lacking symbolic reasoning.

1. Physical Symbol Systems (PSS): The Foundation of Symbolic AI

1.1 Definition and Core Principles

A Physical Symbol System (PSS) is a computational framework where cognition emerges from manipulating symbols (abstract representations) using formal rules (operations). Introduced by Allen Newell and Herbert Simon (1976), PSS posits that:

• Symbols are discrete entities (e.g., words, numbers, or logical expressions) that stand for objects, concepts, or actions.
• Rules define how symbols are combined, transformed, or retrieved (e.g., "IF A AND B, THEN C").
• Physical system: The hardware/software that processes these symbols (e.g., a CPU executing logic gates).

Visual: Symbol Manipulation in a PSS

graph LR
    A["Symbol: CAT"] -->|"Rule 1"| B["Symbol: ANIMAL"]
    B -->|"Rule 2"| C["Symbol: HAS_FOUR_LEGS"]
    C -->|"Rule 3"| D["Symbol: MAMMAL"]
    D -->|"Rule 4"| E["Conclusion: CAT is a MAMMAL"]

1.2 How PSS Works: A Step-by-Step Trace

Consider a simple rule-based expert system for diagnosing car engine problems (inspired by real diagnostic tools like Ncell’s customer support chatbots):

1. Symbolic Knowledge Base:

   • SYMBOL: Engine_Noise
     • RULE: IF Engine_Noise = "Knocking" THEN Check_Oil_Level
   • SYMBOL: Oil_Level
     • RULE: IF Oil_Level = "Low" THEN Replace_Oil

2. Input: User reports "Knocking" noise.

3. Inference Engine applies Rule 1 → triggers Check_Oil_Level.

4. Output: System asks, "Is the oil level low?"

   • If Yes, it concludes: SYMBOL: Problem = "Low_Oil" → SYMBOL: Solution = "Replace_Oil".

Worked Example: Daraz Order Fulfillment Queue Daraz’s order processing system uses a PSS-like pipeline to handle customer orders:

flowchart LR
    A["Customer Order"] --> B["Symbol: ORDER_ID"]
    B --> C["Rule: Validate Payment"]
    C -->|"Success"| D["Symbol: STATUS = 'Paid'"]
    D --> E["Rule: Dispatch to Warehouse"]
    E --> F["Symbol: STATUS = 'Shipped'"]
    F --> G["Rule: Update Tracking"]

• Symbols: ORDER_ID, STATUS, TRACKING_NUMBER.
• Rules: Payment validation, warehouse dispatch, tracking updates.
• Real-world tie: If a rule fails (e.g., payment invalid), Daraz’s system symbolically flags it for manual review, just like a human would.

1.3 Advantages and Limitations of PSS

2. Language of Thought (LOT) Hypothesis

2.1 Definition and Proponents

The Language of Thought (LOT) hypothesis (Jerry Fodor, 1975) argues that:

• The mind represents knowledge using an internal symbolic language (like a mental programming language).
• This language is compositional: Complex thoughts are built from simpler symbolic combinations (e.g., DOG + BARKS → "DOG_BARKS").
• LOT is amodal: It operates independently of sensory modalities (e.g., you can "think" about a cat without seeing/hearing it).

2.2 LOT vs. Physical Symbol Systems

2.3 How LOT Explains Cognition

LOT provides a framework for:

1. Thought Composition: Combining symbols to form new ideas.
   • Example: SYMBOL: APPLE + SYMBOL: RED → THOUGHT: "RED_APPLE".
2. Modular Processing: Different mental modules (e.g., vision, memory) use LOT symbols to communicate.
3. Problem-Solving: Mental "algorithms" manipulate LOT symbols (e.g., solving a math problem).

Visual: LOT Symbol Composition

graph TD
    A["Mental Symbol: APPLE"] --> B["Mental Symbol: RED"]
    A --> C["Mental Symbol: ROUND"]
    B & C --> D["Combined Thought: RED_ROUND_APPLE"]

2.4 Criticisms of LOT

• Symbol Grounding Problem (Harnad, 1990): How do symbols get meaning? LOT assumes symbols are inherently meaningful, but how?
• Neural Realism: The brain doesn’t seem to use discrete symbols; neurons encode distributed patterns (see Unit 8).
• Embodied Cognition: LOT ignores how the body interacts with the environment (e.g., grasping an object requires sensorimotor feedback).

Real-World Example: WhatsApp’s "Read Receipts" WhatsApp uses a symbolic representation of message status:

• SYMBOL: MESSAGE_ID = "123"
• SYMBOL: STATUS = "SENT" → STATUS = "DELIVERED" → STATUS = "READ".
• LOT Connection: Your mental model of "seen" vs. "unseen" messages mirrors this symbolic system.

3. Symbol Grounding and the Embodied Mind

3.1 The Symbol Grounding Problem

PSS and LOT assume symbols have inherent meaning, but:

• Problem: How does a symbol like DOG connect to the real-world concept of a dog?
• Example: A robot told "The dog is under the table" must ground DOG and UNDER in sensor data and motor actions.

3.2 Solutions and Alternatives

4. PSS in AI: Applications and Failures

4.1 Success Stories

1. Expert Systems:

   • eSewa’s Tax Calculator: Uses PSS rules to compute taxes based on income symbols (INCOME_AMOUNT, TAX_BRACKET).
   • Medical Diagnosis: MYCIN (1970s) used PSS to diagnose infections from symptom symbols.

2. Natural Language Processing (NLP):

   • Early chatbots (e.g., ELIZA) matched input symbols to predefined responses.
   • Example:

     flowchart LR
         A["User: I feel sad"] --> B["Symbol: SAD"]
         B --> C["Rule: Respond with Empathy"]
         C --> D["Bot: That sounds tough. Want to talk?"]

3. Game AI:

   • Chess Engines (e.g., Deep Blue): Evaluate board positions using symbolic rules (e.g., PAWN_AT_E4).

4.2 Failures and Limitations

1. Perception Tasks:

   • PSS struggles with raw data (e.g., YouTube’s video recognition uses neural networks, not symbols).

2. Common-Sense Reasoning:

   • PSS lacks world knowledge. Example: A PSS might not infer "The bird is flying" from "The wings are flapping" without explicit rules.

3. Dynamic Environments:

   • Self-Driving Cars: PSS cannot adapt to unpredictable scenarios (e.g., a child suddenly running into the road).

Worked Example: NTC’s Traffic Light System (PSS vs. Reality)

• PSS Approach: Traffic lights use symbolic rules:
  • SYMBOL: CAR_AT_RED_LIGHT → RULE: STOP.
  • SYMBOL: PEDESTRIAN_PRESSED_BUTTON → RULE: CHANGE_TO_WALK.
• Real-World Issue: PSS fails to handle:
  • A drunk driver running a red light (no SYMBOL: DRUNK_DRIVER rule).
  • Emergency vehicles (requires SYMBOL: AMBULANCE + priority rules).

5. PSS vs. Connectionist Models (Neural Networks)

Visual: PSS vs. Neural Network for Digit Recognition

graph LR
    subgraph PSS Approach
        A["Symbol: Digit"] --> B["Rule: Compare to 7 stored shapes"]
        B --> C["Output: 7"]
    end
    subgraph Neural Network
        D["Input: Pixel Grid"] --> E["Hidden Layer: Feature Extraction"]
        E --> F["Output Layer: Probability Distribution"]
        F --> G["Output: 7 (95% confidence)"]
    end

Neural networks excel at recognizing 7 from messy handwriting, while PSS fails without perfect symbol matches.

## In the Real World

1. eSewa’s Tax Calculation:

   • PSS Idea: Uses symbolic rules to compute taxes based on income brackets (SYMBOL: INCOME_1M, SYMBOL: TAX_RATE_10%).
   • How: Input your income → system matches it to predefined tax symbols → outputs tax amount.
   • Limit: Fails if you have irregular income (e.g., freelance earnings).

2. Pathao’s Ride Dispatch System:

   • PSS Idea: Matches SYMBOL: RIDER_REQUEST to SYMBOL: NEAREST_DRIVER using location symbols (LATITUDE, LONGITUDE).
   • How: When you request a ride, Pathao’s system:
     1. Creates SYMBOL: ORDER_ID = "1001".
     2. Applies RULE: Find Driver within 2km.
     3. Updates SYMBOL: STATUS = "Driver Assigned".
   • Real Picture: The "driver assigned" notification is a PSS symbol being grounded in your phone’s UI.

3. Ncell’s Network Diagnostics:

   • PSS Idea: Uses symbolic error codes (e.g., SYMBOL: ERROR_404) to diagnose network issues.
   • How: When your call drops, Ncell’s system:
     1. Logs SYMBOL: CALL_DROP.
     2. Applies RULE: Check Tower Coverage.
     3. If SYMBOL: TOWER_DOWN, it routes you to a working tower.
   • Limit: Cannot predict new types of network failures (e.g., a solar flare disrupting signals).

## Exam Tip

1. Define Clearly:

   • PSS: "A computational system where cognition arises from manipulating symbols via formal rules."
   • LOT: "The hypothesis that mental processes operate using an internal symbolic language."

2. Compare PSS and Neural Networks:

   • Exam Question: "How do Physical Symbol Systems differ from connectionist models in handling perception?"
   • Answer Structure:
     • PSS: Uses discrete symbols; struggles with raw data (e.g., images).
     • Neural Networks: Distributed representations; excels at perception but lacks symbolic reasoning.
     • Example: PSS fails to recognize a handwritten 7; neural nets succeed.

3. Symbol Grounding:

   • Key Point: The challenge of linking symbols to real-world meaning.
   • Exam Tip: Mention Rodney Brooks’ perceptual symbols or situated cognition as solutions.

4. Real-World Applications:

   • eSewa/Khalti: Symbolic rules for transactions.
   • Pathao/Ncell: Symbolic order/diagnostic systems.
   • YouTube/Google: Neural networks for perception (contrast with PSS).

5. Worked Examples:

   • Always show steps (e.g., Daraz order processing trace).
   • Use small numbers (e.g., tax calculation with income = Rs. 500,000).

6. Criticisms:

   • Symbol Grounding Problem: "How do symbols get meaning?"
   • Embodied Cognition: "PSS ignores the body’s role in cognition."

Final Visual Summary

mindmap
  root((Physical Symbol Systems))
    Definition
    Components
      Symbols
      Rules
    Applications
      Expert Systems
      NLP
      Game AI
    Limitations
      Symbol Grounding
      Perception
    Comparison
      vs Neural Networks