Please enable JavaScript.

Coggle requires JavaScript to display documents.

MRTTLN 20AI services, Amazon Q, epochs, có một lưu ý là nếu ta quá sa đà…

- - - - Clustering: group similar data points together into cluster based on their features. Vd. customer segmentation nhóm các khách hàng thành từng nhóm dựa trên các sản phẩm họ đã mua: nhóm người ăn chạy mua rau củ, nhóm người trẻ/ sinh viên mua beer/ chips/ pizza
      - Association Rule Learning: dùng thuật toán Apriori. VD. mua bánh mì thường sẽ mua bơ
      - Anomaly Detection: Isolation Forest (iForest), Vd. fraud detection
      - Dimensionality reduction: reduces the number of features in dataset
    - - Classification: predict categorical label of input data
      - Regression: predict a numeric value based on input data ex. predicting house prices, stock price prediction, weather forescating
      - Training 80%/ Validation 10%/ Test Set 10%
      - Feature Engineering: Feature Extration, Feature Selection, Feature Transformation using domain knowledge to select & transform raw data into meaningful features. Vd. đổi birthDate thành Age giúp tăng performance MLs model
      - Decision tree: splitting the data into subset based on most significiant feature
      - Neural network
      - Dùng trong Fine-tuning
    - - Application: Gaming - teaching AI to play complex game (Chess, Go), Robotics - navigating & manipulating objects in dynamic environments, Finance - portfolio management & trading strategies, Healthcare - optimizing treatment plans, Autonomus Vehicles - path planing & decision making
      - Key concepts: Agent, Environment, Action, Reward, State, Policy
      - RLHF Reinforcement Learning From Human Feedback: RL có cơ chế reward, nên thay vì reward sẵn có mặc định thì dựa vào human feedback để reward
    - - Domain Adaptation Fine-Tuning: tiếp tục huấn luyện model trên tập dữ liệu lớn unlabeled dataset cho các các vụ chung. Khác với Fine-tuning train trên những tập nhỏ đặc thù, labeled dataset cho output là high performance on a specific task
      - Continued Pre-Training: sử dụng tập data của một domain mới như y tế, tài chính, luật.
        Uses unlabeled data to pre-train a model, whereas, fine-tuning uses labeled data to train a model.
- - - - Recall (độ nhạy): best when false negatives are costly
      - F1 Score: best when you want balance between precision & recall, especially in imbalanced datasets
      - Accuracy (độ chính xác tổng thể): best for balanced dataset
      - Precision (độ chính xác): best when false positives are costly
- - - - AMZN Open Search Service: search & analytics db, real time similarity queries, store milions of vector embeddings, scalable index management, fast nearest-neighbor (kNN) search capability
      - AMZN DocumentDB (with MongoDB compatibility): noSQL db
      - AMZ Aurora relational db proprietary on AWS, AMZN RDS for PostgeSQL
      - AMZN Neptune: graph db
    - - Tune: Grid search, random search. SageMaker Automatic Model Tuning (AMT)
    - - Batch: provide discount upto 50%, multiple predictions at a time (output is a single file in S3)
      - Provisioned Throughtput
        
        throughput: max. number of input.output tokens processed per minute
        
        work with Base, Fine-tune, Custom Models
        
        purchase model unit for a certain time 1month, 6months
      - On-Demand
        
        Text Models - charged for ever input/output token processed
        
        Embedding Models - cahrged for everyone input token
        
        Image Models: charged every image generated
        
        Pay-as-you-go (no commitment)
        
        Work with Base Model only
      - Cost saving opt.
        
        Batch: upto 50% discount, wait for result
        
        Provisioned Throughtput: usually not a cost-saving measure, great to "reserve" capacity
        
        On-Demand: unpredictable workload, no commitment
        
        Temperature, Top K, Top P: no impact on pricing
        
        Model size: usually a smaller model will be cheaper (based on providers)
        
        Number of Input/ Output Tokens
  - - - Batch inference: xử lý các batch dữ liệu lớn cùng lúc mà ko y/c phản hồi tức thì
      - Asynchronous inference: mỗi y/c được xử lý riêng biệt, ko y/c trả KQ ngay lập tức, payloads up to 1 GB
      - Serverless inference: tự động quản lý resources
      - Real-time inference: xử lý thời gian thực, chatbot/ recommendation system