AI-900 Azure AI Fundamentals

Key points:

Neural networks learn by adjusting weights
They attempt to emulate human brain learning
The goal is accurate predictions on both training AND new data

Why option B is correct:

When weights are adjusted too precisely to match training data patterns
This leads to "overfitting" - memorizing training data rather than learning general patterns
Results in poor performance on new, unseen data

Why other options are incorrect:

A) Adding more layers might make overfitting worse
C) More computing power won't solve the fundamental issue of overfitting
D) The problem isn't network complexity; it's how the weights have been optimized

The key insight is that while neural networks learn through weight adjustment, too precise optimization to training data prevents the model from generalizing well to new situations - similar to memorizing answers versus understanding concepts.

The Universal Language Model used by the speech-to-text API is optimized for conversational and dictation scenarios. The acoustic, language, and pronunciation scenarios require developing your own model.

Face identification in the Azure AI Face service can address one-to-many matching of one face in an image to a set of faces in a secure repository. Face verification has the capability for one-to-one matching of a face in an image to a single face from a secure repository or a photo to verify whether they are the same individual. Face attributes, the find similar faces operation, and Azure AI Custom Vision do not verify the identity of a face.

Clustering is a machine learning type that analyzes unlabeled data to find similarities present in the data. It then groups (clusters) similar data together. In this example, the company can group online customers based on attributes that include demographic data and shopping behaviors. The company can then recommend new products to those groups of customers who are most likely to be interested in them. Classification and multiclass classification are used to predict categories of data. Regression is a machine learning scenario that is used to predict numeric values.

Weather temperature and weekday or weekend are features that provide a weather temperature for a given day and a value based on whether the day is on a weekend or weekday. These are input variables for the model to help predict the labels for e-scooter battery levels, number of hires, and distance traveled. E-scooter battery levels, number of hires, and distance traveled are numeric labels you are attempting to predict through the machine learning model.

A clustering algorithm is an example of unsupervised learning, which groups data points that have similar characteristics without relying on training and validating label predictions. Supervised learning is a category of learning algorithms that includes regression and classification, but not clustering. Classification and regression algorithms are examples of supervised machine learning.

Linear regression is a machine learning algorithm module used for training regression models. The Clean Missing Data module is part of preparing the data and data transformation process. Select Columns in Dataset is a data transformation component that is used to choose a subset of columns of interest from a dataset. Evaluate model is a component used to measure the accuracy of trained models.

The business card model analyzes and extracts key information from business card images and includes common data field extractions, such as name and email. The invoice model extracts key information from sales invoices and includes common data fields used in invoices for extraction. The read model, layout model, and general document model do not identify and extract common data fields.

Key phrase extraction is used to extract key phrases to identify the main concepts in a text. It enables a company to identify the main talking points from the support question data and allows them to identify common issues. Named entity recognition can identify and categorize entities in unstructured text, such as people, places, organizations, and quantities. The Azure AI Speech service, Conversational Language Understanding, and Azure AI Bot Service are not designed for identifying key phrases or entities.

Knowledge mining is an artificial intelligence (AI) workload that has the purpose of making large amounts of data searchable. While other workloads leverage indexing for faster access to large amounts of data, this is not their primary purpose.

Key points:

Binary classification has two possible outcomes (positive/negative)
The model has high accuracy for negative cases (95% correct)
The model has lower accuracy for positive cases (60% correct)
In medical contexts, missing disease cases (false negatives) can be life-threatening

Why other options are incorrect:

A) Overall accuracy isn't the most important metric in medical diagnosis
B) The model actually performs well with healthy patients (95% correct)
D) Processing time isn't mentioned and isn't relevant to the accuracy issue

The key insight is that in medical applications, missing actual disease cases (false negatives) is typically more dangerous than incorrectly flagging healthy patients (false positives), as it could lead to untreated conditions with serious health consequences.

The Split Data module is particularly useful when you need to separate data into training and testing sets. Use the Split Rows option if you want to divide the data into two parts. You can specify the percentage of data to put in each split, but by default, the data is divided 50-50. You can also randomize the selection of rows in each group, and use stratified sampling.

A webpage or an existing document, such as a text file containing question and answer pairs, can be used to generate a knowledge base. You can also manually enter the knowledge base question-and-answer pairs. You cannot directly use an image or an audio file to import a knowledge base.

Data mining workloads primarily focus on the searching and indexing of data. The computer vision can be used to extract information from images, but it is not a search and indexing solution. Conversational AI is part of natural language processing (NLP) and facilitates the creation of chatbots. Semantic segmentation provides the ability to classify individual pixels in an image depending on the object that they represent.

Key points:

Transformers learn semantic relationships between words
They determine probable sequences based on training data
They use encoder blocks for vocabulary representation
They use decoder blocks for generating sequences

Why option C is correct:

The model only understands statistical relationships between words
It can generate "probable" sequences that make linguistic sense
It has no actual understanding of medical practice or consequences
It's working exactly as designed, but this reveals its fundamental limitation

Why other options are incorrect:

A) More training data wouldn't give the model real medical understanding
B) The encoder is likely working properly if responses are technically correct
D) The architecture isn't the issue; it's the inherent limitation of text prediction

The key insight is that transformer models, despite their sophistication, only process semantic relationships and probabilities - they don't truly understand the real-world implications of the text they generate.

Image generation models can take a prompt, a base image, or both, and create something new. These generative AI models can create both realistic and artistic images, change the layout or style of an image, and create variations of a provided image.

Before you can start training a machine learning model, you must first create a pipeline in the Machine Learning designer. This is followed by adding a dataset, adding training modules, and eventually deploying a service.

Clustering algorithms group data points that have similar characteristics. Regression algorithms are used to predict numeric values. Classification algorithms are used to predict a predefined category to which an input value belongs. Supervised learning is a category of learning algorithms that includes regression and classification, but not clustering.

Each phrase returned by an image description task of the Azure AI Vision includes the confidence score. An endpoint and a key must be provided to access the Azure AI Vision service. Bounding box coordinates are returned by services such as object detection, but not image description.

Identifying potential harms is the first stage when planning a responsible generative AI solution.

Computer Vision: Needed to interpret and understand the handwritten text in the prescriptions (visual interpretation)
Document Intelligence: Required to process and manage these documents in a structured way

Embeddings is an Azure OpenAI model that converts text into numerical vectors for analysis. Embeddings can be used to search, classify, and compare sources of text for similarity.

Tokenization is part of speech synthesis that involves breaking text into individual words such that each word can be assigned phonetic sounds. Transcribing is part of speech recognition, which involves converting speech into a text representation. Key phrase extraction is part of language processing, not speech synthesis. Lemmatization, also known as stemming, is part of language processing, not speech synthesis.

The deep learning models are characterized by their deep architectures, which consist of multiple layers of interconnected neurons. These layers allow the model to learn hierarchical representations of data, where each layer extracts increasingly abstract features from the input data. Training these models often requires processing large amounts of data through numerous layers, leading to longer training times compared to simpler machine learning algorithms.

Option A is INCORRECT because the deep learning algorithms typically require large amounts of training data. The use of data for predictions is not a contributing factor to the longer training times.

Option B is INCORRECT because while this is a characteristic of deep learning, it’s not directly related to longer training times. Deep learning algorithms can create high-level features autonomously, but this doesn’t necessarily lead to longer training times.

Option C is INCORRECT because this option describes a characteristic of machine learning, not deep learning. Deep learning algorithms often tackle problems on an end-to-end basis rather than breaking them into smaller steps.

Option D is INCORRECT because in deep learning, features are typically learned from the data automatically, without requiring user input. This involvement of users in feature engineering is more common in traditional machine learning approaches.

Microsoft recognizes six principles of responsible AI, mentioned below.

Fairness, Reliability and safety, Privacy and security, Transparency, Inclusiveness and Accountability.

All other options are incorrect.

Key points:

Multiclass classification must assign exactly one class to each input
Classes are mutually exclusive (an input can't belong to multiple classes)

Why option B is the biggest challenge:

Mixed sentiment ("The service was great but the price was terrible") can't be accurately captured by a single class
The model must force the feedback into one category, losing important information

Why other options present lesser challenges:

A) Language translation could be added as a preprocessing step
C) Short feedback might be vague but still fits into one category
D) Spelling errors can be handled with text preprocessing

Image classification is part of computer vision and can be used to evaluate images from an X-ray machine to quickly classify specific bone fracture types. This helps improve diagnosis and treatment plans. An image classification model is trained to facilitate the categorizing of the bone fractures. Object detection is used to return identified objects in an image, such as a cat, person, or chair. Conversational AI is used to create intelligent bots that can interact with people by using natural language. Facial detection is used to detect the location of human faces in an image.

When categorizing an image, the Azure AI Vision service supports two specialized domain models: celebrities and landmarks. Image types is an additional capability of the computer vision service, allowing it to detect the type of image, such as a clip art image or a line drawing. Both people_ and people_group are supported categories when performing image classification.

Named Entity Recognition (NER) is a Text Analytics service that helps identify entities in the text and group them into different entity categories, like person, organization, location, event, etc.

You should expect three recognized named entities with the category “Location” in the API response- “headquarters,” “Paris,” and “Eiffel tower.”

Live speech translation involves the following sequence of the services during the real-time audio stream- Speech-to-Text -> Speech Correction -> Machine Translation -> Text-to-Speech.
Options A is incorrect because the Speech Recognition is not involved in this process. You can define from and to translation languages in Speech Translation service settings.
Option E is incorrect because we are not using Text Analysis services in this case.

The invoice model extracts key information from sales invoices and is suitable for extracting information from sales account documents. The ID document model is optimized to analyze and extract key information from US driver’s licenses and international passport biographical pages. The business card model, receipt model, and language model are not suitable to extract information from passports or sales account documents.

Image description is not a capability included in the DALL-E model, therefore, it is not a use case that can be implemented by using DALL-E, while the other three capabilities are offered by DALL-E in Azure OpenAI.

NaN, or not a number, designates an unknown confidence score. Unknown is a value with which the NaN confidence score is associated. The score values range between 0 and 1, with 0 designating the lowest confidence score and 1 designating the highest confidence score.

Key points:

Regression is specifically for predicting numeric values
The model is trained using features and known numeric labels

Why option B is the LEAST aligned:

Predicting a category ("senior" level) is a classification task, not regression
Regression predicts continuous numeric values, not categorical labels

Why other options, despite some issues, still align with regression:

A) Predicting a specific numeric value is exactly what regression does
C) While giving a range isn't typical regression output, it's still dealing with numeric values
D) While negative salary is illogical, it's still a numeric prediction (model might need improvement)

Object detection is the form of ML that helps to recognize objects on the images. Each recognizable object will be put in the bounding box with the class name and probability score

Speech recognition uses audio data to analyze speech and determine recognizable patterns that can be mapped to distinct user voices. Azure AI Speech synthesis is concerned with vocalizing data, usually by converting text to speech. Azure AI Speech translation is concerned with multilanguage translation of speech. Language identification is used to identify languages spoken in audio when compared against a list of supported languages.

Embeddings is an Azure OpenAI model that converts text into numerical vectors for analysis. Embeddings can be used to search, classify, and compare sources of text for similarity.

In a regression machine learning algorithm, features are used to generate predictions for the label, which is compared to the actual label value. There is no direct comparison of features or labels between the validation and training datasets.

Object detection can be used to evaluate traffic monitoring images to quickly classify specific vehicle types, such as car, bus, or cyclist. Linear regression is a machine learning training algorithm for training regression models. Image classification is part of computer vision that is concerned with the primary contents of an image. OCR is used to extract text and handwriting from images.

After we ingest the data, we need to do a data preparation or transformation before supplying it for model training. There are four typical steps for data transformation such as Feature selection, Finding and removing data outliers, Impute missing values, and Normalize numeric features.

Option C is incorrect because Split data is coming after data transformation.
Option D is incorrect because ML algorithm selection data is coming after data transformation and Split Data steps.

For language model training, we need to provide the following key elements- Entities, Intents, and Utterance. We can achieve this by using the Azure Cognitive service LUIS portal.

Entity is the word or phrase that is the focus of the utterance, as the word “light” in the utterance “Turn the lights on.”

Intent is the action or task that the user wants to execute. It reflects in utterance as a goal or purpose. We can define intent as “TurnOn” in the utterance “Turn the lights on.”

Utterance is the user’s input that your model needs to interpret, like “Turn the lights on” or “Turn on the lights”.

All other options are incorrect.

Tagging involves associating an image with metadata that summarizes the attributes of the image. Detecting image types involves identifying clip art images or line drawings. Content organization involves identifying people or objects in photos and organizing them based on the identification. Categorizing involves associating the contents of an image with a limited set of categories.

To deploy a predictive service from a newly trained model by using the Machine Learning designer, you must first create a pipeline in the Machine Learning designer. Adding training modules by using the Machine Learning designer takes place before creating a trained model, which already exists. Adding a dataset by using the Machine Learning designer requires that a pipeline already exists. To create an inferencing cluster, you must use Machine Learning studio.

Key points:

The process is sequential
Each stage builds on the previous one
Missing something in Stage 1 affects all subsequent stages

Why option A is correct:

Stage 1 (Identify potential harms) is foundational
If reputational harm wasn't identified as a potential issue:
It wouldn't be measured in Stage 2
No mitigation strategies would be planned in Stage 3
No operational guidelines would address it in Stage 4

Why other options are less critical oversights:

B) You can't properly measure what you haven't identified as a potential harm
C) You can't mitigate what you haven't identified and measured
D) Operational guidelines can't address unidentified issues

The key insight is that failing at Stage 1 creates a cascade of problems throughout the entire process. If you don't identify a potential harm at the beginning, you can't properly address it in any subsequent stage.