Useful Tips to Prepare for the Azure AI 900 Exam

So you want to pass the AI-900 Exam?

Start by remembering these handy tips! 🏆

Image classification is a common computer vision task to search through images. Image classification helps to classify images based on image content.

Object detection helps to identify objects and their boundaries.

Semantic segmentation helps classify pixels to the objects they belong.

Face detection is a computer vision technique that helps detect and recognise people's faces.

Image analysis helps extract information from images, tag them and create descriptive image summaries.

Computer vision cognitive services include:

·       Detect objects

·       Categorise image

·       Identify landmarks

·       Read the text in an image

To create a language model, we need 3 essential elements in our data:

·       Entities

·       Utterances

·       Intents

We can do this in Azure Cognitive LUIS portal.

Utterance = the input that the model needs to interpret.

Entity = the focus word ‘light’ in the phrase ‘turn on the light’

Intent = the action ‘turn on’ in the phrase ‘turn on the light’

In evaluating regression models Azure machine learning uses:

·       Relative absolute error RAE

·       Mean absolute error MAE

·       Coefficient of determination R-squared/R2

·       Root mean squared error RMSE

·       Relative squared error RSA

In a custom vision model, you get object detection information including:

·       Probability score %

·       Class name

·       Bounding box

Three main authoring tools on Azure Machine Learning Studio home page:

·       Automated ML

·       Notebooks

·       Designer

Four typical steps of data transformation:

·       Impute missing values

·       Normalise numeric features

·       Find and remove outliers

·       Feature selection

Typical Azure ML cycle – we ingest data, then we transform it, then we split it to train and validate, then we select the machine learning algorithm, then we train it, then we test it and evaluate it.

Split data comes after transformation. Machine learning algorithm selection comes after split.

AI services on Azure has a QnA Maker that allows you to create a knowledge base from an excel xlsx file.

Types of compute resources on Azure ML Studio:

·       Compute clusters

·       Kubernetes clusters

·       Compute instances

·       Attached compute

A classification model can use historic data to predict a category that a new item belongs to based on labels (category1/2/3/4/5 etc.)

Azure Bot Service can create a virtual assistant and connect it to various input channels and devices.

Three metrics used to evaluate custom vision performance:

·       Precision

·       Average Precision (AP)

·       Recall

Form Recognizer supports custom model & pre-built receipt model.

Custom model is your own data and requires 5 samples to start.

Pre-built receipt model is a form recognizer trained to work with receipts.

The six Principles of Responsible AI:

·       Fairness

·       Reliability and safety

·       Privacy and security

·       Inclusiveness

·       Transparency

·       Accountability

Resistance to harmful manipulation is an example of AI principle ‘reliability and safety’

‘accountability’ directs AI solutions to follow governance and organisational norms

Positive sentiment in sentiment analysis is 1.0, neutral is 0.5, negative 0.0, and sentiment analysis is a common scenario of NLP workloads.

Feature engineering is creating new features.

Normalising numeric features to a common scale is part of data prep and transformation. Normalising numeric features is not feature engineering as it is not creating new features.

Data mining focuses on searching and indexing data.

Regression predicts a numeric value in the future based on historic data.

In regression a validation set contains both known features and known labels.

To train a machine learning model in machine learning designer you first need to create a pipeline.

Object detection provides bounding box coordinates to identify different objects within an image.

Face identification is 1 to many.

Face verification is 1 to 1.

Face identification and Face verification are found in Azure AI Face Vision Studio.

Stemming normalises words.

Tokenization breaks text into individual words.

AI Language Service in Azure provides language section features which returns:

·       ISO code 6391

·       Language name

·       Score

Azure AI Services provides direct access to AI translator and AI speech. Azure AI services provide direct access to both AI translator and AI speech through a single endpoint and authentication key.

System messages should be used to set the context for the model by describing expectations.

Embeddings is an Azure open AI model that can be used to search classify and compare text for similarity.

Semantic segmentation provides the ability to classify individual pixels in an image to the object that they represent.

Azure AI vision includes OCR and Spatial analysis.

Three parts of the machine learning process taken care of by Azure AI vision include:

·       Evaluating a model

·       Training a model

·       Choosing a model

Two typical AI workload scenarios for NLP are translating text between languages and performing sentiment analysis.

Removing stop words is the first step in the statistical analysis of terms used in a text in the context of NLP.

Azure AI Speech has an NLP workload used to generate closed captions for live presentations.

Azure AI Speech provides speech to text and text to speech through speech recognition and speech synthesis.

Good luck in your exam! 👍🤞🔥

More exam questions here 👈