Overview

Disclaimer: These are my personal learning notes. Do not consider them an official source of truth.

The AI-900 certification provides foundational knowledge of artificial intelligence (AI), including concepts, methods, and evaluation techniques across various AI workloads and models.

Vision

Computer vision enables AI systems to interpret and process visual information from images or videos.

Image Classification: Identifies objects within an image.
Object Detection: Identifies objects and their locations within an image.
Semantic Segmentation: Identifies boundaries and segments individual objects within an image.

Speech

Speech technologies enable AI systems to understand and generate human speech.

Speech Recognition: Converts spoken language to text.
Speech Synthesis: Converts text into spoken language.

Natural Language Processing (NLP)

NLP allows AI systems to comprehend, analyze, and generate human language.

Entity Extraction: Identifies named entities (people, places, dates, etc.) within text.
Text Classification: Assigns documents or text into predefined categories.
Sentiment Analysis: Determines the emotional tone (positive, negative, neutral) within the text.
Language Detection: Identifies the language in which a text is written.

Note: NLP can also be referred to as Natural Language Understanding (NLU).

Supervised Learning

Supervised learning involves training models on datasets containing labeled features, allowing the model to predict outcomes based on learned patterns.

Classification

Used when labels are categorical.

Binary Classification

Labels belong to one of two possible categories.
Common algorithm: Logistic Regression.

Multiclass Classification

Labels belong to one of three or more possible categories.

Common algorithms:

One-vs-Rest (OvR): Trains a binary classifier for each category separately, then combines the results.
Multinomial Classification: Predicts multiple categories, providing a probability for each category.

Evaluation Metrics

Evaluated using a confusion matrix and related metrics:

Accuracy:
- Accuracy can be misleading if the data is imbalanced. For example, 95% accuracy on a dataset with 95% negatives means the algorithm always returns False.
Recall:
- Measures how many actual positives were correctly identified.
Precision:
- Measures how many predicted positives were correct.

Regression

Regression is used when predicting continuous numerical outcomes based on input data.

Evaluation Metrics

Mean Absolute Error (MAE): Average absolute difference between actual and predicted values.
- Treats all errors equally regardless of direction.
Mean Squared Error (MSE): Average squared difference between actual and predicted values.
- Penalizes larger errors more significantly than smaller ones.
Root Mean Squared Error (RMSE): Square root of MSE; measures error in the same unit as the label.
Coefficient of Determination (R²): Indicates how much of the variance in the dependent variable can be explained by the model.
- Ranges from 0 (poor fit) to 1 (perfect fit).

Unsupervised Learning

Unsupervised learning involves training models on datasets without labels, discovering patterns or structures in the data based solely on features.

The most common unsupervised technique is clustering, grouping data points based on similarity.

K-means clustering: Groups data into K clusters by finding the nearest neighbors and iteratively adjusting cluster centroids.

Evaluation Metrics

Evaluating unsupervised models can be challenging due to the absence of labels.

Silhouette Score: Measures how similar a data point is to its own cluster compared to others.
- Ranges from -1 to 1; higher scores indicate better-defined clusters.