Deep Learning and Related Topics - Summary
Applications of Deep Learning
Used in image recognition, NLP, healthcare (diagnosis), finance (fraud detection), and autonomous vehicles.
Historical Milestones
1957: Perceptron
1986: Backpropagation
2012: AlexNet revolution
2014+: GANs, ResNet, Transformer
Machine Learning vs Deep Learning
ML: Needs feature engineering, works with less data.
DL: Learns features automatically, needs more data.
Forward Propagation
Input -> Weighted sum -> Activation -> Output
Backpropagation Steps
1. Forward pass
2. Compute loss
3. Backward pass (gradients)
4. Update weights
Feedforward Neural Network Requirements
Labeled data, Differentiable activations, Proper weight initialization, Optimizer
Activation Functions
ReLU, Sigmoid, Tanh, Softmax
Fine-Tuning
Deep Learning and Related Topics - Summary
Use a pre-trained model on a new task by retraining the final layers (e.g., ResNet on medical images).
CNN in Image Classification
Extract features using convolution -> pool -> flatten -> classify.
Transfer Learning Applications
Medical imaging, NLP, object detection, speech, sentiment analysis.
Deep Q-Learning
Learns policies using Q-values; good for high-dimensional spaces but training is slow.
Reinforcement Learning Components
Agent, Environment, Policy, Reward, Value function
Q-Learning
Updates Q-values using rewards and estimated future values.
Challenges in Deep RL
Sparse rewards, training time, stability, exploration
Cross-Validation
Split data into K folds, rotate training/testing, average results.
Precision, Recall, F1-Score
Precision: Accuracy of positives
Recall: Coverage of positives
F1: Balance between them
Deep Learning and Related Topics - Summary
Confusion Matrix
Shows TP, FP, TN, FN - used to evaluate classification models.
K-Fold Use Cases
Medical diagnosis, NLP, finance models, image classification
Object Detection in CNN
Detect objects and their locations using models like YOLO, SSD.
Sentiment Analysis (NLP)
Classifies text as positive/negative using embeddings + deep models.
Medical Image DL Model
Use CNN on X-rays or MRIs to classify or segment abnormalities.