Please enable JavaScript.
Coggle requires JavaScript to display documents.
machine learning overview - Coggle Diagram
machine learning overview
Training Supervision
: how they are
supervised
?
Supervised
learning
training data include the desired solutions:
labels
examples:
classification
Unsupervised
learning
training data is
unlabeled
examples
dimensionality reduction
simplify the data without lose much information
anomaly detection
association rule learning
clustering
Self-supervised
learning
Involves generating a labeled dataset from a fully unlabeled one.
Semi-supervised
learning
Combination of supervised and unsupervised techniques
Reinforcement
learning
Agent
observes the environment
perform actions
get rewards in return
must learn by it self what is the best strategy
(policy)
Environment
Batch versus Online Learning:
they can learn
incrementally
on the fly?
Batch
learning
incapable of learn incrementally
it must be trained using all the
available data
take a
lot of time
. require lots of
computing resources
Offline learning
: it's trained first and them apply what it has learned
model rot
or
data drift
: when model performances decay slowly over time
Online
learning
train the system
incrementally
by feeding it data instances sequentially
data can be
individually
or
small groups
(mini-batches)
a challenge is that if bad data is fed to the system, the performance will decline
learning rate
: parameter that tells the systems how fast they should adapt to changing data
if a
high learning rate
was set, the system will rapidly adapt to new data, but tend to forgot the old data
low learning rate
means more inertia (learn more slowly). good thing is: will be less sensitive to noise and outliers
How they generalize?
Instance
-based versus
Model
-based learnin
Instance-based
learning
the system learn the examples by heart, then generalizes
similarity measures
to compare
Model-based
learning
create a model from a set of examples
use the model to make predictions
Overall
Model selection
: chose the type of model and fully specified its architecture
Training a model
: running an algorithm to find the model parameters that will make it best fit the training data