3. Deep Learning, Neural Network and Machine Learning

Machine Learning

Supervised Learning

• Classification

• Regression

Unsupervised Learning

• Clustering

Supervised Learning

• Labelled datasets

• High accuracy and reliability

• Data Security

• Costly to aquire data

• Poor ability to extrapolate predictions outside of sample data given

• Vulnerable to biases in labelled data

• can try semi supervised learning

Unsupervised Learning

• Group and interpret data based on input data’s characteristics

• No labelling cost

• High flexibility

• Hard to quantify model’s abilities

• Uncertain outcomes

• Computationaly Complex

Reinforcement Learning

Rewarding desired behaviours and punishing undesired ones. No labels. Learns through interactions with an environment. Learns from consequences. 

Reinforcement Learning Vs Unsupervised and Supervised Learning

You should use Reinforcement Learning (RL) over Supervised or Unsupervised Learning when the problem involves sequential, real-time decision-making in a dynamic environment where there are no clear historical examples of the "perfect" solution.

Neural Network

• Mimick the neural network of our brains

• Learns by repeatedly adjusting internal weights so that its outputs get closer to the desired results.

• Input, Output and Hidden layers

• Backward Propagation and Forward propagation

Neural Network Definitions

• Input layer: receives the raw data and passes it into the network.

• Hidden layers: perform computations with weights and activation functions to learn patterns.

• Output layer: produces the final prediction (class, probability, value).

• Forward propagation: data flows from input → hidden → output to generate a prediction.

• Backward propagation: the error is calculated and sent backward to adjust weights so the model improves

RNN / LSTM

• Used for tasks like sentence generation and machine translation

• Short term memory due to vanshing and exploding gradients

• LSTM utilises input, output and forget gates along with memory cell states to enable long term memory

CNN

• Used for image processing via convolution and pooling

• Uses convolution kernel layers to detect edges and shapes

  • Pooling downsamples the kernel keeping only the important feature, reducing computation required

AI Usecase Summary

Machine Learning

• Recommendation, Predictions & Detections

Neural Networks

• Basic Classification/Regression

Deep Neural Networks

• Vision, Language, Gen AI