Page 34 - CyberSecurityEssentialsEnglish
P. 34
TABLE 1-4: COMMON WAYS MODELS LEARN
What does the
Learning method Simple example Example use
system learn from?
Learning from
Examples linked to
examples with Messages labeled Classify messages
correct answers or
answers (supervised “spam/not spam.” or requests.
labels.
learning)
Finding patterns
without given Unlabeled data in
Unlabeled usage Group customers
answers which the system
records. by similar behavior.
(unsupervised looks for similarities.
learning)
Learning from parts of One part of a text is Predict a hidden Train language
the data itself (self- used to predict part or the next models on large
supervised learning) another. token. text collections.
Learning through
Feedback indicating
experience and Learn a game Games or control
whether the
reward strategy or control over sequential
system’s choice
(reinforcement policy. steps.
succeeded.
learning)
Can the Model Succeed on New Examples?
We do not want a model merely to memorize its training examples, a problem known as
overfitting. We want it to generalize, that is, to succeed on new cases that resemble the
environment in which it will operate. A model that performs well on its training examples
but poorly on new ones will not be useful, no matter how high its training score.
Developers therefore test a model on examples that were not used in training. A person
evaluating an existing tool can likewise prepare a local test set of realistic cases, then
compare the results against a clear standard before adopting the tool.
What Happens When the Data Changes?
A strong model can lose performance when its data or operating environment changes, a
situation known as dataset shift. A model trained on formal messages may make mistakes
with short conversations full of dialect and abbreviations. A demand-forecasting model may
weaken after an economic or seasonal change absent from its training data. Performance
must therefore be monitored after deployment, rather than relying only on initial test results
[12].
33
