Tag: DataScience

Rules for What We Know, AI for What We Don’t 

Posted on by lechuck park

This image presents a practical guide on how to effectively integrate Artificial Intelligence, specifically Large Language Models (LLMs), into software systems. The overarching theme is “Rules for What We Know, AI for What We Don’t,” which emphasizes using reliable, traditional computing for hard facts and reserving AI for complex reasoning and interpretation.

1. Don’t Prompt What You Can Query

This principle warns against using AI to retrieve exact data. Because LLMs generate responses based on probabilities, they can sometimes guess incorrectly or hallucinate. If you need a verified fact—like a user’s bank balance—you should use a standard database search to fetch that exact number. Once you have the accurate data, you can then pass it to the AI to draft a natural, polite response.

2. Connect the Certain, Compute the Complex

This section suggests building a hybrid approach to problem-solving. You should establish a strict, rule-based foundation (the “certain”) using traditional logic, math, or physics. Once that solid framework is in place, you let the AI operate on top of it to handle creative or flexible tasks (the “complex”). For example, use traditional software to ensure a building is structurally safe, and then use AI to design creative interior layouts within those safe boundaries.

3. LLM is the Engine, Not the Database

This final point clarifies the true role of an LLM: it is a processor, not a storage drive. You shouldn’t try to force an AI to memorize massive amounts of raw data, like a 10,000-page company manual. Instead, use a search system to find the exact page you need, and then feed just that relevant text into the LLM. The AI acts as the “engine” to read, understand, and summarize that specific information for you.

Summary

To build reliable AI applications, rely on traditional databases and strict logic for factual retrieval and structural constraints. Use LLMs strictly as reasoning and processing engines to interpret context, draft text, and solve complex problems based on the hard facts you provide them.

#AIArchitecture #LLM #ArtificialIntelligence #SoftwareEngineering #DataScience #PromptEngineering #GenerativeAI

PIML(Physics-Informed Machine Learning)

Posted on by lechuck park

PIML (Physics-Informed Machine Learning) Explained

This diagram illustrates how PIML (Physics-Informed Machine Learning) combines the strengths of physics-based models and data-driven machine learning to create a more powerful and reliable approach.

1. Top: Physics (White-box Model)

  • Definition: These are models where the underlying principles are fully explained by mathematical equations, such as Computational Fluid Dynamics (CFD) or thermodynamic simulations.
  • Characteristics:
    • High Precision: They are very accurate because they are based on fundamental physical laws.
    • High Resource Cost: They are computationally intensive, requiring significant processing power and time.
    • Lack of Real-time Processing: Complex simulations are difficult to use for real-time prediction or control.

2. Middle: Machine Learning (Black-box Model)

  • Definition: These models rely solely on large amounts of training data to find correlations and make predictions, without using underlying physical principles.
  • Characteristics:
    • Data-dependent: Their performance depends heavily on the quality and quantity of the data they are trained on.
    • Edge-case Risks: In situations not covered by the data (edge cases), they can make illogical predictions that violate physical laws.
    • Hard to Validate: It is difficult to understand their internal workings, making it challenging to verify the reliability of their results.

3. Bottom: Physics-Informed Machine Learning (Grey-box Approach)

  • Definition: This approach integrates the knowledge of physical laws (equations) into a machine learning model as mathematical constraints, combining the best of both worlds.
  • Benefits:
    • Overcome Cold Start Problem: By using existing knowledge like mathematical constraints, PIML can function even when training data is scarce, effectively addressing the initial (“Cold Start”) state.
    • High Efficiency: Instead of learning physics from scratch, the ML model focuses on learning only the residuals (real-world deviations) between the physics-based model and actual data. This makes learning faster and more efficient with less data.
    • Safety Guardrails: The integrated physics framework acts as a set of safety guardrails, providing constraints that prevent the model from making physically impossible predictions (“Hallucinations”) and bounding errors to ensure safety.

#AI #PIML #MachineLearning #Physics #HybridAI #DataScience #ExplainableAI #XAI #ComputationalPhysics #Simulation

with Gemini