Microsoft Azure AI Engineer Associate Certification Concept Review Notes

This is a quick 5 min recap of 43 AI-102 Learning Path modules to review before taking the certification exam:

I. Plan and manage an Azure AI solution (15–20%) Section

• Diagnostic settings captures data for subsequent analysis.
• You must specify API Key, Billing, and Eula parameter with the value “yes” when deploying an Azure AI services container.
• Alert Rules include scope (resource you want to monitor), condition (signal type or metric), optional actions, and alert rule details.
• Microsoft’s Responsible AI principles include: Fairness, Reliability and Safety, Privacy and Security, Inclusiveness, Transparency, and Accountability.
• Azure AI services are spread across 5 categories: Natural Langauge Processing, Knowledge mining and document intelligence, Computer vision, Decision support, and Generative AI.

II. Implement decision support solutions (10–15%)

• Text moderation classification has a score between 0 and 1 to indicate the predicted category for the evaluated text; sexually explicit content is category 1, sexually suggestive is category 2, and offensive language is category 3.
• Azure AI Personalizer uses reinforcement learning that chooses the best action for a given context, aiming to maximize a reward.
• The exploration value in Personalizer can find new patterns and change based on new discoveries instead of using the trained model’s prediction.
• Azure AI Personalizer can learn and mimic your app’s current state when set on Apprentice mode. Once the Reward Achievement Ratio reaches 75–85% (meaning the ratio of the average reward of the Personalizer over the average reward of your app’s default logic), you can change it to Online mode to pursue best actions.
• Personalizer’s inference explainability adds feature scores which helps you find out why a particular recommendation was made to your users and any biases your model might have by identifying features that have the most/least influence on the model’s decisions.
• When debugging or testing Azure AI Personalizer, set update frequency to 1 minute as recommended.

III. Implement computer vision solutions (15–20%)

• To publish a trained Azure AI Custom Vision model, you can host it either on an Azure AI Custom Vision (Prediction) resource or an Azure AI Services resource.
• The main difference between image classication model and object detection model is whether it’s tagging the whole image or each object (region) in an image.
• The Face service enables you to: Detection, Attribute analysis*, landmark location, comparison, recognition, liveness.

*head pose, glasses, blur, exposure, occlusion, accessories, QualityForRecognition

• Optical Character Recognition (OCR) extracts text from images.
• For receipts, articles, and invoices, use Document Intelligence. For street signs, handwritten notes, and store signs, use Image Analysis.
• Before using Azure Video Indexer to extract information from a video, index it first.

IV. Implement natural language processing solutions (30–35%)

• For Azure language detection, the document size must be under 5,120 characters.
• If the text we’re trying to analyze is ‘I saw Venus shining in the sky’, extracting linked entities will identify ‘Venus’ along with a Wikipedia page about Venus.
• To improve performance of a knowledge base, you can use active learning (create question & answer pairs/review suggestions) and define synonyms.
• Email addresses, home addresses, IP addresses, names, and protected health information are Personally Identifiable Information (PII).
• Conversational language understanding (CLU) is a conversational AI service that builds custom natural language understanding models after the user teaches it how to predict intents and entities.
• “What time is it?” “What is the time?” “Tell me the time” are all utterances that a user might enter while intent is what the user wants to perform — GetTime.
• If you want to get the results of a classification task via the API, you should call the URL provided in the ‘operation-location’ header in the request response.
• Precision and recall are two metrics that check if the entity recognition works. Recall indicates how well the model extracts entities, while precision is how the correct label is applied to extracted entities.
• Converting “こんにちは” to “konnichiwa” is an example of transliteration, which renders the text into a different script rather than a translation to a language.
• The Azure AI Speech SDK requires the location and a key to connect to the Azure AI Speech service.
• Speech to speech translations are called synthesized translations: You can use event-based synthesis to translate from one source language into a single target language.

V. Implement knowledge mining and document intelligence solutions (10–15%)

• Pricing tiers include: (1) Free, (2) Basic — small scale search solutions (3) Standard — enterprise scale solutions (4) Storage Optimized
• Depending on the pricing tier, you can optimize for scalability and availability by creating replicas and partitions: Replicas (R) are instances of the search service, while Partitions (P) are used to divide an index into multiple storage locations. The number of search units is R x P = SU.
• Lexical analysis is one of the Query Processing stages when query terms are refined based on rules such as lowercase conversion, removal of stopwords, and root form conversion (ex: comfortable -> comfort).
• By default, search results are sorted by a relevance score based on a term-frequency/inverse-document-frequency (TF/IDF) algorithm, which measures the relevance of a word/corpus to the text.
• To ensure the files can be accessed by Azure AI Search in a blob container, a data source should be created that allows files to be searchable.
• A skillset enables you to define an enrichment pipeline that integrates AI skills into the indexing process.
• After creating an index, you need to apply ‘retrievable’ attribute to the index field to ensure that it can be included in search results.
• After implementing a custom skill as an Azure function, if you want to include the custom skill in Azure AI Search indexing process, you need to add a WebApiSkill to a skillset with the URI for the function.
• To enrich an index with different language translations, you can use Azure AI services; Azure Speech Service is focused on translating speech to text and text to speech services.
• For Azure AI Machine Learning Studio models, the endpoint has to be an Azure Kubernetes Service (AKS) as container instances aren’t supported.
• The best way to improve the performance of an Azure Machine Learning (AML) skill when enriching documents is to improve the performance of your inference cluster by using more powerful nodes in the Kubernetes inference cluster.
• 503 HTTP response status is when request is rejected due to heavy load and 207 HTTP response status is when some documents failed in the batch.
• To improve the availability of search solution is (1) to increase the number of replicas (2) to use the Availability Zones.
• Common search solution issues (search latency, throttled search percentage, delete search service, stop search service) can be notified using alerts.
• Semantic search aims to improve the ranking of search results, which uses the language understanding model(s) with the BM25* ranking function.

*ranking results based on the frequency of that term within a document.

• Semantic captions are extracted summary sentences from the document verbatim that highlights the most relevant text.
• Embeddings are vector representations of semantic meaning of a text that allows machine learning models to find semantic similarity.
• To map values to projects in a knowledge store, you can use the Shaper skill to create a simpler JSON structure than documents built iteratively by the skills in the enrichment pipeline.
• A composed model takes different types of forms into a single service, identifies and selects the most appropriate custom model to use for analysis.
• Only custom models that have been trained with labeled example forms can be added to a composed model so if you receive an error, you should check that the custom models were trained with labels.
• Key phrase extraction, language detection, merge, sentiment, translation, image analysis, and optical character recognition are all skills that call to an AI process that enriches the index.
• Azure Machine Learning (AML) custom skills and Custom Web API skills are two types of custom skill.
• In document cracking, the indexer opens the content files and extracts their content and in field mappings, fields are extracted from the content.

VI. Implement generative AI solutions (10–15%)

• For Completion playground, there are 8 parameters: (1) Temperature controls randomness (2) Max length sets a limit on tokens per model response (3) Stop sequences (4) Top probabilities controls randomness by choosing model’s token selection to likelier tokens (5) Frequency penalty (6) Presence penalty (7) Pre-response text (8) Post-response text

*It is recommended to change either temperature or top_p at a time, but not both.

• Chat in Azure OpenAI Studio playground is able to support conversation-in, message-out scenarios.
• Completion, ChatCompletion, and Embeddings are the three available endpoints for interacting with a deployed Azure OpenAI model.
• Fine-tuning is when you create a custom model by training an existing model (ex: gpt-35-turbo) with a dataset of additional training data. Pros: higher quality responses Cons: costly and time intensive.
• Breaking down the task and using chain of thought prompting can help Azure OpenAI model respond more effectively.
• Microsoft’s Responsible generative AI guidance has four stages: Identify potential harms, Measure harm in the outputs generated, Mitigate the harms at multiple layers, Operate the solution responsibly.
• To mitigate potential harm in generative AI solutions, four layers can be included: (1) The model layer (2) The Safety system layer (3) The metaprompt and grounding layer (4) The user experience layer.

Photo by Sunrise King on Unsplash

If you’re renewing your certification, you can refer to my previous article ‘5 Tips for Renewing Your Microsoft Azure AI Engineer Certification as a Non-Engineer’.