Please enable JavaScript.

Coggle requires JavaScript to display documents.

MRTTLN MLA-C01, Xem sâu vô toán phát là kinh hoàng, không hiểu gì hết luôn…

- - - - Can replicate DELETE markers from source to target (optional). Delete with a version ID are not replicate
      - No "chainning" of replication
      - Only new object are replicated after you enable this feature. Using S3 Batch Replicate to replicate existing object.
  - - - Performance Mode
      - Throughput Mode
        
        Elastic: unpredictable IO
        
        Bursting
        
        Enhanced
        
        Provisioned
      - EFS Scale
      - Storage Classes
        
        Infrequent Access (EFS-IA)
        
        Archive
        
        Standard
      - Availability & durability
        
        Standard: multi AZ
        
        One Zone: EFS One Zone-IA
- - - - Tabular Data: dự đoán doanh số, phân loại KH & các bài toán dự báo
        
        CatBoost: xử lý dữ liệu table có nhiều biến phân loại. High performance open-source implementation of the Gradient Boosting Decision Tree (GBDT) algorithm. GBDT is a superviesed learning.
        
        Linear Learner: dùng cho phân loại và hồi quy tuyến tính
        
        Auto-Gluon Tabular: tự động xây dựng model autoML. Open source autoML fw that trains highly accurate ML models on an unprocessed tabular dataset
        
        K-Nearest Neighbors (k-NN): phân loại và hồi quy bằng cách tìm các điểm gần nhất. Index-based algorithm, it use non-parametric for classification or regression.
        
        XGBoost (Extreme Gradien Boosting): thuật toán cây quyết định tăng cường phổ biến. Work well with imbalance dataset, adjust the 'scale_pos_weight' parameter which allows you to give more importance to the minority class (fraudulent transaction).
        Prallel processing, it can handle large datasets and complex models without execssive computational costs, making it suitable for real-time fraud dection system
        
        Factorization Machines algorithm is a general-purpose supervised learning that you can use for both classification and regression tasks
        
        LightGBM: popular & efficient open-source implementation of GBDT algorithm
        
        TabTransformer is a novel deep tabular data modeling architecture for supervised learning. https://docs.aws.amazon.com/sagemaker/latest/dg/tabtransformer.html
        
        AdaBoost (Adaptive Boosting), focusing on correcting the errors of weak classifiers, assigning more weight to misclassified in each iteration
      - Unsupervised Learning
        
        Principle Component Analysis (PCA) demisionality reduction
        
        K-Mean group similar object/ data together
        
        Random Cut Forest (RCF) anomaly detection in real-time
        
        Latent Dirichlet Allowcation (LDA), Neural Topic Model (NTM) organize a set of docs into topic
        
        IP Insight IP anomaly detection
      - Textual analysis
        
        Blazing Text algo, Text Classification TensorFlow: assign pre-defined categories to docs in a corpus, ctegorize books in a library into academic disciplines
        
        Sequence-to-Sequence algo: machine translation, text summarization, speech to text
      - Supervised Learning
        
        Predict if an item belong to a category: email SPAM filter.
        
        Predict a numeric/ continuos value: estimate the value of the house: XGBoost, CatBoost, k-NN, LightGBM, Linear Learner, Factorization Machine, Tab Transformer
        
        Based on historical dat for a behavior, predict future behavior: DeepAR algo
        
        Improve the data embeddings of a high dimensional objects: Object2Vec algo
      - Image processing
        
        Object detaction - MXNext, TensorFlow: detect people & obj in an IMG
        
        Semantic Segmentation algo: self driving car prepare to identify object in their way
        
        Image classification, MXNext: alerts about adult content in an IMG
      - Ensemble Learning
        
        Bagging training the same model on different substs of the data
        
        Boosting each model trying to correct the error of the previous one
        
        Staking combine multiple models, leverage their strength while minimizing their weakness
        
        Simple Voting may not result in a significant improvement if the models have very different performances or if the base models are weak
    - - Tuning
        
        Random Search: select groups of hyperparameters randomly on each iteration. Its work well when a relative small number of the hyperparameters primarily determine the model outcome
        
        Grid Search: tedious & computionally intensive, especially with a large of hyperparameters. Less efficient for Byesian Optimization for complex model.
        
        Bayesian Optimization: the algo builds a probailistic model from a set of hyperparameters that optimizes a specific metric. Its uses regeression analysis to interatively choose the best set of hyperparameters
      - SageMaker automactic hyperparametr tuning
        
        Training algo
        
        Performance metric
    - - Ground Truth: label data tự động hoặc manual với con người
      - Clarity: Có 21 chỉ số, giảm thiểu bias. Feature enginerring, evaluate the feature importance, help u identify the most impactful features and futher refine the model
        
        https://docs.aws.amazon.com/sagemaker/latest/dg/clarify-data-bias-metric-cddl.html
        
        Post-training data & model bias metric
        
        Pre-training Bias metric
        
        Class Imbalance (CI)
        
        Label Impalance (DPL)
        
        Jensen-Shannon Divergence (JS)
        
        Lp-norm (LP)
        
        Kullback-Leibler Divergence (KL)
        
        Kolmogorov-Smirnov (KS)
        
        Conditional Demographic Disparity (CDD)
        
        Total Variation Distance (TVD)
      - Canvas: no-code ML design tool for benginner & manager.
      - Feature Store: central repository tập hợp các features
      - Data Wrangler: công cụ xử lý dữ liệu trước khi train model. DynamoDB is not natively supported=> Use Boto3
      - Model Dashboard: theo dõi hiệu suất và trạng thái các model ML đã triển khai
      - JumpStart: triển khai nhanh các model ML và giải pháp AI dựa trên template sẵn có. You can evaluate, compare and select FMs quickly based on pre-defined quality and responsibility metric
      - Model cards: documents critical detail of ML model for governance & auditing
      - MLFlow: kết hợp cùng SageMaker để quản lý vòng đời của các model triển khai trên SageMaker. Manage machine learning experiments
      - Amazon SageMaker Debugger
      - Pipeline: similar to MLFlow, its automates the process of building, training, deploying ML model. Model Lineage: track the history of model deployment, Model Group organizes versions of models, Model Package represent a version of a model, Model Metric store model performance
      - Script Mode buil ML models using custom code and algo. Often use with "Bring Your Own Container" (BYOC) approach ECR (Elastic Container Registry)
      - Model Registry manage lifecycle of ML models: track & manage different version of model, store metadata such as tranining data/ hyperparameter/ evaluation metrics, define custom stage dev/ tesing/ prod, share model with other user, automate model deployment with CICD
      - Model Monitor monitor the quality of the ML models, address data drift, to track changes in input data distribution. Ensures that models perform as expected and alerts users to issues that might require human intervention.
      - Autopilot build models you can almost use in production, you can use when you have missing data, provides stastical insight about column in your dataset.
        Automates the process of selecting algo and tuning hyperparameters.
    - - Batch inference: xử lý các batch dữ liệu lớn cùng lúc mà ko y/c phản hồi tức thì. Get prediction for an entire dataset
      - 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
      - SageMaker Edge Manager: edge device like smartcamera, robots, personal computer, mobile devices
      - Optimize performane using Neo: Gluon, Keras, MXNet, PyTorch, TensorFlow, ONNIX
    - - On demand
      - Managed Spot Training: allowcate spot instance, saving upto 90% cost
    - - S3 Mode File
      - File input mode with EFS use case require transferring data from S3 to EFS
      - S3 Pipe input mode data pre-fetched from S3 at high concurrency & throughput, streamed into the named pipe (FIFO pipe)
      - S3 FastFile Input Mode: download entire dataset to the training instance before starting the training job => large dataset would leads to signification delays & require large amounts of local storage, which is not optimal for efficiency or cost. Only using with S3!
    - - Schedule Scaling
      - On-Demand Scaling
      - Step Scaling Policies define specific threadsholds
      - Target Tracking Scaling Policies AUTOMATICALLY adjust the number of instance in response to change in demand by tracking a specific metric
- - - - Precision (độ chính xác): best when false positives are costly
      - Recall (độ nhạy): best when false negatives are costly
      - F1 Score: best when you want balance between precision & recall, especially in imbalanced datasets. Models predictive performance.
      - Accuracy (độ chính xác tổng thể): best for balanced dataset
  - - - L2 Reguarization (Ride)
      - L1 Regularization (Lasso)