Machine Learning

• Machine learning is prerequisite for Deep Learning
• It is split into Supervised Learning and Unsupervised Learning

ML consists of: #

• Task (Image Classification, Assigning labels) -> If only two kinds of labels then “Binary classification”
• Experience (Data)

Other types of Machine Learning #

Reinforcement learning #

• Have an agent that can interact with an environment
• Environment will give reward (positive reinforcement)

Self-supervised lerning #

• Like supervised learning but no manual labour is necessary
• For example future frame prediction in videos
• The model gets the correct answer itself

How does learning work #

• Important difference between memorizing and learning (“channelize”)
• For learning: Partition data into two sets:
  • Training data
  • Test data

-> To prevent memorizing

• Training and Test data should have same distributions of attributes
  • Example: The training data and the test data for autonomous driving might be totally different. The available footage rarely shows tricky situations and thus might not prepare the model for such situations.

Measuring the learning success #

• After training the model with the training data:
• How well does the model do on the test data
• Possible Performace Measure:
  • Accuracy: “How many elements were classified correctly?”

How to obtain the model #

1. Use training data and some random weights
2. Get some predictions out of it
3. Compare the predictions with the ground truth using a Loss Function
4. Optimize the weights based on the Loss Function