• Contents:

Linear Learner

Ref: https://www.udemy.com/course/aws-certified-machine-learning-engineer-associate-mla-c01/learn/lecture/45285403

• 🔧 Fit a line into dataset
  • Linear regression → predictions based on linear function
  • Classification → uses linear function as threshold
    • Binary or multi-class function
  • ‼️ Linear learner can do much more than linear regression!
    • Works similarly to DL neural NWs
  • 💡 If a dataset doesn't look like it can fit a line well, better use another algorithm
  • Diagram
• Example: classify handwritten numbers from pixels

Input Data (Preprocessing)

• File & Pipe modes supported
• Valid formats:
  • RecordIO-Protobuf (Float32 data only)
    • most performant
  • CSV (First column = label)
• Should be normalized
  • Either yourself or tell Linear Learner to do it
  • Why? → So all features are weighted equally
• Should be shuffled → gets better results

Training

• Uses Stochastic Gradient Descend (SGD)
  • Many optimization algorithms supported (SGD, Adam, Adagrad…)
• Trains multiple models in parallel
  • Most optimal model chosen during validation
• Offers L1 & L2 regularization to prevent overfitting
• Instance Types
  • single or multi-machine CPU
  • single-machine GPU
    • ❗ Multi-GPU doesn't help!

Important Hyperparameters

• Balance_multiclass_weights
  • each class gets equal importance in loss functions
• Learning_rate & mini_batch_size