Google Cloud Computing Foundation Course - Week 13 Lecture Notes Summary

Building Bespoke Machine Learning Models

• AI Platform Overview: Google’s AI platform helps developers, data scientists, and engineers
streamline the transition from ideation to production and deployment of ML models.
• TensorFlow Basics:
• Open-source library for numerical computation, primarily used for ML but adaptable for
any computational task.
• Uses Directed Acyclic Graph (DAG) to represent computations where nodes perform
operations and edges represent data (tensors).
• Tensors: Multi-dimensional arrays representing data in TensorFlow. Data flows through the
graph, hence the name "TensorFlow".
• Distributed Training:
• Large datasets are often divided into batches and trained across multiple machines.
Scaling out (horizontally) is the solution rather than scaling up (more GPUs).
• Feature Engineering: Human insights can refine data features, helping models perform better.
Preprocessing of raw data is essential.
• Model Deployment:
• After training, models must handle a high number of queries per second (QPS) for
predictions. The TensorFlow API allows interaction with the model from various
languages.
• The AI platform simplifies handling of the complexities of scaling, model versioning,
and dealing with training-serving skew.
• AI Platform Benefits:
• Manages the preprocessing and feature engineering steps, distributes models for high
QPS, and provides prediction through a simple REST API.

Cloud AutoML
• Introduction: Cloud AutoML is a suite of ML products that allows users with minimal ML
expertise to train custom, high-quality models tailored to specific business needs.
• Simplified ML Process: Cloud AutoML reduces the complexity of model building, requiring
less training data and no infrastructure management.
• AutoML Products:
• AutoML Vision: Trains image classification models through an intuitive graphical
interface.
• AutoML Video Intelligence: Trains models to classify and track objects in videos.
• AutoML Natural Language: Enables the training of custom models for sentiment
detection, entity recognition, and classification.
• AutoML Translation: Allows domain-specific translation models to be trained on
custom data.
• AutoML Tables: Transforms raw data into production-ready models, useful in industries
like retail, insurance, and marketing.
Google's Pre-trained Machine Learning APIs
• Pre-trained Models: Provide off-the-shelf solutions for common ML use cases, eliminating the
need to curate datasets or build models from scratch.
• Cloud Vision API:
• Image Classification: Detects entities (e.g., car, cat) and facial features in images.
• Optical Character Recognition (OCR): Extracts text from images, supporting multiple
languages.
• Landmark and Logo Detection: Identifies famous structures and product logos.
• Cloud Speech-to-Text API: Converts real-time or pre-recorded audio into text, supporting over
120 languages.
• Cloud Text-to-Speech API: Converts text into natural, human-like speech with over 180
voices.
• Cloud Translation API: Translates text between languages and detects the source language.
• Cloud Natural Language API: Performs sentiment analysis, entity recognition, and syntax
analysis to understand text.
• Cloud Video Intelligence API: Annotates videos by detecting objects, events, and activities,
making video content searchable and discoverable.

Questions with Answers

1. Q: What is the AI Platform, and how does it help developers?
A: The AI Platform helps ML developers, data scientists, and engineers transition their ML
models from ideation to production with managed services.
2. Q: What is TensorFlow?
A: TensorFlow is an open-source library for numerical computation, primarily used for building
machine learning models.
3. Q: What is a Directed Acyclic Graph (DAG) in TensorFlow?
A: A DAG is a computational graph where nodes represent operations and edges represent data
flow (tensors).
4. Q: What are tensors in TensorFlow?
A: Tensors are multi-dimensional arrays of data used in TensorFlow computations.
5. Q: Why is distributed training important in machine learning?
A: Distributed training allows large datasets to be split into batches and processed across
multiple machines, enabling scalability.
6. Q: What is feature engineering in machine learning?
A: Feature engineering involves refining or adding new data features to improve model
performance based on human insights.
7. Q: How does TensorFlow handle multi-device execution?
A: TensorFlow can execute models on both CPUs and GPUs, providing hardware portability.
8. Q: What is the role of TensorFlow API in model deployment?
A: TensorFlow API allows models to be invoked from any programming language, making it
easier to deploy and scale predictions.
9. Q: What is training-serving skew?
A: Training-serving skew occurs when the data used for training differs from the data used in
 predictions, affecting model performance.
10.Q: What benefits does the AI Platform offer for model deployment?
A: The AI Platform manages scaling, versioning, and prediction requests, simplifying model
deployment.
11.Q: What is Cloud AutoML?
A: Cloud AutoML is a suite of products that allows users with limited ML expertise to train
high-quality custom models with minimal effort.
12.Q: How does Cloud AutoML simplify machine learning?
A: It eliminates the need to manage infrastructure, code models, or use extensive training data,
allowing faster model development.
13.Q: What is AutoML Vision used for?
A: AutoML Vision allows users to train image classification models through a graphical
interface.
14.Q: What is the AutoML Natural Language API used for?
A: It is used for training models that detect sentiment, classify entities, and extract information
from text.
15.Q: What industries can benefit from AutoML Tables?
A: Industries like retail, insurance, and marketing can use AutoML Tables for predictive
analytics on customer behavior, demand forecasting, and fraud detection.
16.Q: What is a pre-trained model in Google Cloud?
A: Pre-trained models are ML models provided by Google, ready for use without the need for
training data or model building.
17.Q: What is the Cloud Vision API used for?
A: The Cloud Vision API is used for image classification, facial detection, OCR, and identifying
landmarks and logos.
18.Q: How does the Cloud Speech-to-Text API help businesses?
A: It converts real-time or recorded audio into text, useful for transcribing meetings, voice
commands, or call center interactions.
19.Q: What does the Cloud Text-to-Speech API do?
A: It converts text into natural, human-like speech, ideal for customer service, device
interactions, and more.
20.Q: What is the Cloud Translation API?
A: It is a service that translates text between languages, with the ability to detect the source
language.
21.Q: What is the Cloud Natural Language API used for?
A: It performs tasks like syntax analysis, sentiment analysis, and entity recognition to
understand the meaning of text.
22.Q: How does the Cloud Video Intelligence API work?
A: It annotates video content by identifying objects and actions, making video more searchable
and accessible.
23.Q: What is TensorFlow's main computational concept?
A: TensorFlow uses Directed Acyclic Graphs (DAGs) for computational tasks, where nodes
represent operations and edges represent data.
24.Q: What is the significance of using the AI Platform for machine learning?
A: The AI Platform manages preprocessing, training, and deployment, reducing the
complexities involved in building and scaling ML models.
25.Q: What are tensors?
A: Tensors are multi-dimensional arrays that represent the data in TensorFlow’s computation.
26.Q: How does feature engineering improve machine learning models?
A: It refines or adds new features based on human knowledge, improving model performance.
27.Q: What is the difference between AutoML Vision and AutoML Video Intelligence?
A: AutoML Vision classifies images, while AutoML Video Intelligence detects and tracks
objects in videos.
28.Q: What does the Cloud Translation API do?
A: It translates text from one language to another and can detect the source language if
unknown.
29.Q: What is the main advantage of pre-trained models?
A: Pre-trained models eliminate the need to collect data and train models from scratch,
providing instant predictions.
30.Q: What is the purpose of hyperparameter tuning?
A: It adjusts the settings of a model during training to improve its performance and accuracy.
31.Q: How does the AI Platform handle prediction requests?
A: It provides both online (real-time) and batch prediction services through a REST API.
32.Q: How does TensorFlow handle execution across devices?
A: TensorFlow allows the same Python code to run on both CPUs and GPUs, ensuring
flexibility in execution.
33.Q: What is AutoML Translation?
A: AutoML Translation allows custom models to be trained for specific language pairs for more
accurate translations.
34.Q: What is the Cloud Natural Language API used for?
A: It analyzes text for sentiment, syntax, and entity recognition to understand the text better.
35.Q: How does Cloud AutoML simplify model building?
A: It automates the process, eliminating the need for coding and infrastructure management.
36.Q: What are the advantages of using the AI Platform for scaling ML models?
A: The AI Platform automatically scales models to handle large volumes of prediction requests
without manual intervention.
37.Q: How does the Vision API handle OCR?
A: It extracts text from images and converts it into searchable, selectable content.
38.Q: What industries can benefit from Cloud AutoML?
A: Retail, marketing, healthcare, and finance, among others, can use AutoML for predictive
modeling and personalized customer experiences.
39.Q: What is the main advantage of pre-trained models over custom models?
A: Pre-trained models are ready for use, reducing the need for time-consuming data collection
and training.
40.Q: What are the main components of TensorFlow’s architecture?
A: The key components are tensors (data), operations (nodes), and edges (data flow).