M12 - Unsupervised Learning Flashcards
What are the categories of machine learning?
- Supervised learning
- Unsupervised learning
- Reinforcement learning
Explain Supervised Learning and give examples.
- labeled data
- feedback
EX. Regression, SVM, Neural Networks
Object recognition
Explain Unsupervised learning and give examples.
- no labeled data
- no feedback
Ex. K-mean, self-organizing maps
Clustering
Explain Reinforcement learning and examples.
- reward based learning
- increase cumulative discounted rewards
Ex. Q-learning, SARSA, TD-learning, DQNs
Robot Navigation
What is the motivation behind using unsupervised learning?
- no hand-labeling data required but we can still perform classification
- find structure (or clusters) in the dataset based on the idea: similar inputs should be located close to each other
- do not need backprop to adjust the weight
What kind of models can we create with unsupervised learning?
- find structures (clusters) in the data
- detect anomalies
- create auto-association
- reduce the dimensionality
When should we use unsupervised learning?
- when we have a lot of sensory data and we can not hand-label them
- we do not have a rewards function to guide the behavior of robot
- we can not structure a score (loss) function to train the network
How does k-means work?
- Partitions data into k number of mutually exclusive clusters
- how well a point fits into a cluster is determined by the distance from that point to the cluster’s centre
What are the convergence criteria for k-means?
- Number of specific iterations
- Until the centres moves with a threshold value
How do we find correct numbers of K?
- Visualize data
- Domain knowledge
Run for different number of K then select the best one
What is competitive learning?
- paradigm different that supervised learning and reinforcement learning
- the aim is not to map input and output by minimizing a cost function
- no explicit reward function to tune the behaviour of network
- form a unit (or neuron) which can be activated for similar inputs
How does competitive learning work?
- Similar inputs will activate the units that are close to each other
- the activation of the neuron will increase if the distance between the neuron weights and input vector is closer
What is the are the Pseudo-code steps of competitive learning?
- Initialize the weights for each unit
- Derive Euclidean distance for each unit for an input
- Select the unit which has the smallest Euclidean distance as the winner unit
- Update the weights of the winner unit
What happens if the learning rate = 0?
No updates will be made to the model’s parameters during the training process
What if the learning rate = 1?
Model’s parameters will be updated by a large step at each iteration
What does competitive learning allow you to do?
- cluster the input
- reduce the dimensions
- extract features in a unsupervised way
What is SOM?
Self organizing map = is a variation of competitive learning
What does SOM do?
- the units close to each other (by weights) respond to similar inputs and have similar weights
What are the limitations of SOM?
- the initialization of the weights may lead to significant differences in the final formation of the SOM
- training duration increases with dimensions of the network
- different SOMs can find different similarities for the same inputs
Competitive learning vs. Self-organizing maps
Competitive Learning:
- Identifies features/prototypes through neuron competition.
- Focuses on individual feature identification.
- Used for feature selection and identification.
Self-Organizing Maps (SOMs):
- Creates a low-dimensional representation of data while preserving topology.
- Emphasizes cluster formation and visualization.
- Used for clustering and visualization tasks.
What are the use cases of SOM’s?
- Epigenetic Robot Architecture (ERA)
- Self-Organized Internal Models Architecture (SOIMA)
Explain Epigenetic Robot architecture.
- the units of the architecture are self-organizing maps
- each SOM can be used to process different modality input
- missing modality representation can be constructed
Explain Self-Organized Internal Models Archtectur.
- the unit of the architecture are self-organizing maps
- implementation of inverse and forward models in the same architecture
- missing modality representation can be constructed
- sensory-motor mapping
