You are on page 1of 19

Artificial Intelligence and Machine

Learning Primer
• Machine Learning
Outline • Gradient Descent and Neural Networks
• Basic Concepts Behind AI models
• Deep Learning Architectures
• Reinforcement Learning
Machine
Learning
• Any process/software that
enables a computer to
learn from data.
Making machines learn:
Gradient Descent Algorithm
• The optimized function is typically the
difference between the target and predicted
values.
Designing the machine’s
brain: Neural Networks
• A graph of consecutive linear
calculations that tries to mimic
the connectivity between the
neurons of the brain
Common Terms and Concepts behind AI models

• Loss Function
• Defined beforehand to calculate how the model
prediction matches the target output.
• Examples:
• Mean Absolute Error
• Mean Squared Error
• Mean Squared Logarithmic Error
• Cross-Entropy or Logarithmic Loss
• Kullback Leibler Divergence
Common Terms and Concepts behind AI models

• Optimizer
• Algorithm that facilitates how the
parameters in the model change as the
training progresses
• Training
• The process wherein the network
parameters (weights) are adjusted to
make the prediction agree as much as
possible to the expected output
• Testing
• Measure the model’s capability to
predict using data not used for training
• Did model learn correctly or just
memorized the data
Deep Learning
Architectures
Convolutional Neural Networks (CNN)

• Mainly used for computed


vision problems such as
object detection and
segmentation
Convolutional Neural
Networks (CNN)

YOLO architecture (Object


Detection)
• You-Only-Look-Once

Mask RCNN (Instance


Segmentation)
• Mask Recurrent-Convolutional-NN
Recurrent Neural Networks

• Mainly used in modelling


sequenced data: Sentences,
voice and music synthesis
and natural language
processing
Recurrent Neural Networks
Neural Parametric Singing Synthesizer (NPSS)
Autoencoders

• a type of artificial neural network used


to learn efficient data codings in an
unsupervised manner.
Variational • VAE applied to Music
Autoencoders
Generative Adversarial Networks

• Two neural networks contest with


each other in a game (in the form of
a zero-sum game, where one
agent's gain is another agent's loss)
Generative Adversarial Networks
Reinforcement
Learning
Reinforcement
Learning
The model is trained for the
purpose of reaching a goal.
Applications of RL

You might also like