18. Cloud-Based AI

Machine Learning Pipeline

Problem Formulation

Buisiness Problem -> ML problem

• ML model: start to think about the problem in terms of your ML model
• Questions
  • Can machine learning solve the problem?
  • Would a traditional approach make more sense?
  • Is t his problem a supervised or unsupervised machine learning problem?
  • Do you have labeled data to train a supervised model?
• Problem + Intended outcome + Make it measurable

 

Collect and evaluate data

Questions

• Which data do you need
• Do you have access to that data?
• How much data do you have and where is it?
• What solution can you use to bring all this data into one centralized repository?

Data Sources

• Private Data
  • Data that you have in various existing systems
  • Data is found in many different system
• Commercial Data
  • Data that a commercial entity collected and made available
• Open-Source Data
  • Comprise many different open-source datasets
  • usually available for use in research or for teaching purposes
  • you can find open-source datasets hosted by AWS, Kaggle ...

 

Data Consideration

• Get an understanding of your data
• Get a domain expert
• Evalulate the quality of your data
  • good data contains a signal about the phenomenon that you're trying to model
• Identify features and labels
  • data = features + labels
  • feature: attribute that can be used to help identify patterns and predict future answers
  • label: answer that you want your model to predict
  • data for which you already know the answer is called labeled data
• Identify labeled data needs
  • supervised learning include labeled data

Data Preparation and Preprocessing

• extract data from one or more data sources
• you need a subject matter expert or a functional expert to understand the authenticity of the data

 

Feature Engineering

Feature Engineering

• dealing with your data to make it usable
• selecting or creating the features

Feature Selection

• selecting the features that are most relevant and discarding the rest
• prevent either redundancy or irrelevance in the existing features or to get a limited number of features to prevent overfitting
• Selection methods
  • Wrapper methods: measure the usefulness of a subset of features by training a model on it and measuring the success of the model
  • Filter methods: use statistical methods to measure the relevance of features, faster and cheaper than wrapper methods because they don't involve training the models repeatedly
  • Embedded methods: algorithm-specific and might use a combination of both

Feature Extraction

• building up valuable information from raw data by reformatting, combining, and transforming primary features into new ones
• encoding data, finding missing data, handling outliers

Preparing Data

• data might be dirty -> contains error or duplicate data
  • using dirty data will lead to incorrect results or an inability to run the model effectively
• Considerations for preparing your data
  • Encoding data
    • ML algo work best with numerical data
  • Cleaning data
    • before you encode the string data, must make sure that the strings are all consistent
    • adjust variables to use a consistent scale
    • split items that capture more than one variable
  • Finding missing data
    • most ML algorithms can't deal with missing values automatically
    • find missing data and update with something meaningful and relevant to the problem
  • Handling outliers
    • outliers: points lie at an abnormal distance from other values
    • outliers can add richness to your dataset but they can also make it more difficult to make accurate predictions
    • outliers affect accuracy because they skew values away from the other more normal values that are related to the feature

• Use feature extraction to convert data into a usable form
• During feature extraction, you will handle msising data, duplicates, inconsistencies, invalid values, and conversion of text data into numerical data

Select and Train model

Model Training

• you don't use all the data to train your model
  • Training data: feeds into the algorithm to produce your model; the model is then used to make predictions over the validation dataset
  • Validation data: you might notice things that you will want to tweak, tune, and change
  • Test data

Benefit of splitting all available data into training(80%), validation(10%), and test(10%) subsets

• The model can be evaluated on data wasn't used for training to assess how well it is generalizing new information

 

Overfitting and Underfitting

• The goal of machine learning is to build a model that generalizes well

What is happening if a model is overfitting?

• The model performs well on the training data, but it doesn't perform well on the evaluation data

 

Chossing a ML algorithm

• Supervised learning (지도학습)
• Unsupervised learning (비지도학습)
• Reinforcement learning (강화학습)

 

Tune and Evaluate model

Evaluate and Tune

• validation data provided a biased evaluation of a model fit while tuning model
  • 모델이 validation data에만 맞출 수 있도록 튜닝될 수 있음
• test data has known values and lets you assess the unbiased accuracy of the model after tuning
  • if the model performs well on the test data, it will likely perform well on new data with unknown target values
• you should set aside enough test data and not use that that data for training or validation

Success Meteric

• Evaluation relies on an appropriate metric
• Model metric should be linked to that business metric as closely as possible

Tuning

• goal of training is a model that is balanced and generalizes well
• modify the model's data, features, or hyperparameters until you find the model that yields the best result

Deploy model

---

Machinne Learning Stack

Data Layer

• where data that will feed directly into your ML model is stored

Model Layer

• contains the model and algorithms that build predictions, based on data that the model collects

Deployment & Monitoring Layer

• model is working in a live environment and producing ML tasks

---

Machine Learning Tools

Tools for Machine Learning

Tools for Machine Learning

• Jupyter Notebook: open-source web application
• JupyterLab: web-based interactive development environment for Jupyter Notebooks
• Pandas: open-source python library used for data handling and analysis
• Matplotlib: generate plots of your data, Python library
• NumPy: scientific computing packages in Python
• Scikit-learn: open-source ML library supports supervised/unsupervised learning

Machine Learning Framework

• provide tools and code libraries that you can use
• AWS supports the following frameworks and they can be used from Amazon SageMaker
  • PyTorch,Caffe2, Torch, TensorFlow, Keras, ApacheMXNet, etc..

 

Amazon Instances Designed for ML

AWS provides compute instances that are tuned for ML, both in the cloud and at the edge

• EC2 C5 instance: cost-effective high performance at a low price for running advanced compute-intensive workloads
  • help to speed up typical ML operations
• EC2 P3 instance: speed up ML applications
  • fastest in the cloud for ML training
  • ideal for ML workloads that need massive parallel processing power
• AWS IoT Greengrass makes it easy to bring intelligence to edge devices
• AWS Elastic Inference is used to attach low-cost GPU-powered acceleration to Amazon EC2 and SageMaker instances

 

AWS offers many managed services that don't require any ML experience

• Computer Vision
  • Amazon Rekognition: provides object and facial recognition for both image and video
  • Amazon Textract: extract txt from recgonition images
• Chat bots
  • Amazon Lex: helps you build interactive conversational apps that use voice or text
• Speech
  • Amazon Polly: converts txt to audible speech
  • Amazon Transcribe: converts spoken audio to text
• Forecasitng
  • Amazon Forecast: use ML to combine time series data with additional variables to build forecasts
• Language
  • Amazon Comprehend: use natural language processing to find insights and relationships in text
  • Amazon Translate: translate text into different langauges
• Recommendations
  • Amazon Personalize: help you create individual personalized recommendations for customers

 

Amazon SageMaker

Amazon SageMaker

• AWS ML service with many capabilities
• help data scientists and developers to Prepare > Build > Train & Tune > Deploy & Manage
• it also provides the ability for developers to iterate the process until they get their model just right
• eveyr computer interfacing step of ML pipeline can be conducted in this one service
• first fully integrated development environment that is designed spcifically for ML
• brings everything that you need for ML under one unified, visual UI

 

 

SageMaker Features

• provides tools for: labeling data, building models, training models, hosting trained models
• can deploy ML nstances that run Jupyter notebooks and JupyterLabs
• In SageMaker, the VMs that you create are your fully managed Amazon SageMaker ML instances
• Data visualization
• Model selection: built-in algorithm, write a script, AWS Marketplace, your own algorithm
• Deployment: you can deploy a model by using a couple of different methods
  • you can use the first way to get one prediction at a time
  • you can use the second way to get predictions on an entire dataset
• Marketplace integration: provide a large selection of ready-to-se model packages and algorithms from ML developers