NotesApplications of Deep Learning
Computer Vision
Can be applied for tasks like image classification, object detection, and image segmentation.
Internet of Things (IoT)
Often used for activity recognition, social sensing, and affective computing.
Speech and Language Modeling
Text to speech and speech to text.
Generative Models
Text to image, large language models.
Fundamental Machine Learning Algorithms
Machine Learning usually encompasses supervised learning, unsupervised learning, and reinforcement learning.
Supervised Learning
Two big classes of problems within this: Classification and Regression.
Classification
Classification predicts integers/classes. The objective is to guess the right class. Essentially, try to take some data and categorize it.
Regression
Regression predicts real values. The objective is to get close to the actual value. Essentially, try to take some set of inputs and guess what the value of those inputs may result in (think linear regression).
Supervised Classification and Regression Tasks
Maps input data to an output prediction.
Classification tasks predict a discrete category/label. E.g. predict main object in an image.
Regression tasks predict a continuous set of values.
What is Inside the Model?
Real world input → Model input → Model → Model output → Real world output
A machine learning model is a mathematical equation described by a set of parameters that determine the input-output relationship. A learning algorithm uses examples of input-output pairs to find a set of parameters (train the model) which predict the relationship most accurately.
Unsupervised Learning
Two big classes of problems within unsupervised learning: Clustering and Association.
Reinforcement Learning
Mainly its own field.
Building off of Linear Models
Linear models can separate non-linear data if we transform the features
 |
|---|
| Source |
| However manual feature engineering becomes complex in high dimensions. E.g. an MNIST image is if we treat it as a vector. |
As a result, we want to find if we can somehow automate or learn this linearization.