NPTEL

Video Course on Machine Learning

Professor Carl Gustaf Jansson, KTH

Week 2 Characterization of
Learning Problems

Video 2.4 Scenarios for

Concept Learning
Basic Distinctions: Supervised vs
Unsupervised Learning

In Supervised learning, input data is always pre-classified with unique

concept labels.The goal of supervised learning is to, based on example input-
output pairs, learn a concept definition which best approximates the
relationship between input data and the concept labels. An optimal scenario
will allow for the algorithm to correctly determine the concept labels for
unseen data-items.

In Unsupervised learning, input data are NOT classified, i.e. contains only
input data and lack concept labels. Unsupervised learning algorithms
therefore have to identify commonalities and structures in the data-set and to
group the input based on similarity. Unsupervised learning algorithms have
to decide on a optimal portfolio of concepts that best matches the data-set and
arrange groupings of subsets of the data-set so that it matches the portfolio of
concepts.
Basic Distinctions:
Off-line (Batch) vs On-line (Incremental) learning
This distinction is relevant for supervised as well as for un-supervised
learning.

Offline learning refers to situations where the system is not operating in a real time
environment but handles pre-harvested data in static and complete batch form.
Most traditional machine learning algorithms are well adapted to off-line learning
and the parallell access to the whole data-set gives full flexibility of using data-
items in all kinds of variations during the learning process.

Online learning is a learning scenario where data is processed in-real time in an

incremental fashion. Input data items are incrementally, gradually and continuously
used to inductively extend the existing model. Results of earlier learning are
typically maintained as still being valid. Incremental algorithms are frequently
applied to data streams or big data. Stock trend prediction and user profiling are
some examples of data streams where new data becomes continuously available.
Many traditional machine learning algorithms inherently support incremental
learning, but may have to be adapted to facilitate this.

A middle way is to handle data in so called mini-batches.

Scenario 1
Learning a single concept off-line from
pre-classified positive examples
Label 1

Concept
Conc 1
Scenario 2 The presence of Noise
Learning a single concept off-line from
pre-classified positive examples
Noise
Label 1
Noise is a fundamental underlying
phenomenon that is present in all datasets.
Noise is a distortion in data, that is unwanted
by the perceiver.

Noise is anything that is spurious and

extraneous to the true data and typically due
to faulty capturing process.
Concept
Conc 1
Noise can occur in all subsequent scenarios
as well as in this first simplest one.
Scenario 3 Outliers
Learning a single concept off-line from
pre-classified positive examples
Outlier
Label 1
An outlier is a data-item that is distant from
other observations.

An outlier may be due to natural but extreme

variation or it may indicate an experimental
error or other noise. An outlier can cause
serious problems for analysis. It is crucial to
distinguish between the measurement error
cases and the cases where the population has
Concept
Conc 1 a heavy-tailed or skewed distribution.

Outliers can occur in all subsequent scenarios

as well as in this first simplest one.
Scenario 4 Negative examples
Learning a single concept off-line from
pre-classified positive examples

- Label 1 Negative examples

For many situations and many machine

+ - learning algorithms, faster convergence

towards a concept definition can be
- achieved by using a mix of positive and
Negative examples.
-
- - Negative examples can be made
available in a variety of ways, either by
Concept
Conc 1
using examples from other labelled
categories or by artificial generation
guided by available domain knowledge.
Scenario 5 Near Misses
Learning a single concept off-line from
pre-classified positive examples
Near Misses
Label 1
- In a scenario where negative examples are
used, it is not obvious, what type of objects we
- - should use as negative examples. Arbitrary
+ negative examples will differ considerably
from positive examples, which might allow the
- -
learner unwanted flexibility in determining the
classification boundary.
-
Near misses are negative examples that differ
from the learned concept in only a small
Concept
Conc 1
number of significant points.

Such examples do not necessarily belong to

known concepts.
Scenario 6 Internal Structure and Topology of Data-set
Learning a single concept off-line from
pre-classified positive examples
Internal Structure and Topology of Data-set
Label 1
In earlier scenarios we have assumed no internal
structure of the Data-set. All data-items have been
P P regarded as having the same status and importance
P and no structure or metric has been assumed among
P the data-items in the set.
P
P In contrast, this scenario introduce the concepts of
P typicality of objects and similarity metrics within the
data set:
- naively more typical objects are more
advantageous to use early in a learning process
Concept
Conc 1
- naively it makes sense to use the similarity metrics
to guide the order of considering training examples.
Scenario 7 Instance-based Learning
Learning a single concept off-line from
pre-classified positive examples
Instance-based learning
Label 1 The 6th scenario, which implies a well defined
Structure and Similarity Metric for the Data-set, opens
up for a new Scenario where no longer an explicit
P P generalization (concept definition) is needed. By
P instance-based learning (memory-based learning) we
P mean learning algorithms that, instead of creating
P
P explicit generalizations, compare new problem
P instances with instances which have already been
stored in memory.
One advantage that instance-based learning has over
other methods of machine learning is its ability to more
easily adapt its model to previously unseen data.
Instance-based learners may simply store a new
instance or may also throw old instances away.
Scenario 8 On-line learning
Learning a single concept on-line from
pre-classified positive examples
On-line learning
Label 1
Online learning is a learning
scenario where data is processed in-
real time in an incremental fashion.
Input data items are incrementally,
gradually and continuously used to
inductively extend the existing
model.

Results of earlier learning are

Concept
Conc 1 typically maintained as still being
valid.
Scenario 9 Throwing away data-items that has become irrelevant
Learning a single concept on-line from
pre-classified positive examples
In the case of On-line learning normally
all data-items that has been encountered
Label 1 and analyzed will be kept resulting in
a monotonically growing Data-set.

However there many reasons why ´older

´ data items may become irrelevant. This
means of course that also the Concept
definitions may have to be revised due
to the´retractions´.

Machine Learning Algorithms that

Concept
Conc 1
handle the On-line case must also be
able to handle this kind of situation.
Scenario 10 This scenario does not introduce many any
big differences and surprises.

All aspects introduced in scenario 1-9 are still

Learning multiple concepts off-line from relevant to consider but in parallell:
- Noise
Pre-classified positive examples - Outliers
- Negative examples
Label 1 Label 2 Label 3 - Near misses
- Internal structure and metrics of data sets
- Instance-based learning
- On-line learning
- Throwing away data-items that has
become irrelevant.

Concept
Conc 1 Concept 2 Concept 3
Scenario 11 Unsupervised concept learning
Learning multiple concepts from
unsorted examples
All aspects introduced in scenario 1-10
are still relevant to consider.

Input data are NOT classified, i.e.

contains only input data observables and
lack output data observables (concept
labels).

Unsupervised learning algorithms

therefore have to identify commonalities
and structures in the data-set and to
group the input based on similarity.

The main category of techniques that

tackles the unsupervised case is called
Cluster Analysis.
Cluster Analysis

Cluster analysis is the assignment of a set of

observations into subsets (called clusters) so that
observations within the same cluster are similar
according to one or more pre-designated criteria, while
observations drawn from different clusters are dissimilar.

Important aspects of clustering techniques are:

- similarity metrics
- internal compactness or density of clusters
- degree of separation - the difference between
clusters.
Clustering analysis can potentially learn also Concept Hierarchies
Category 1
Generalization
Generalization
Category 2 Category 3 Category 4

Generalization Generalization

Category 5 Category 6 Category 7 Category 8 Category 9 Category 10

The end-to-end process for Concept Learning

In a typical Concept Learning task the following steps need to be considered:

• Data harvesting from potentially heterogeneous sources

• Pre-processing of data (e.g. from analogue to digital form)

• Establishment of model or theory support

• Feature engineering

• Algorithm selection

• Tailoring conditions for algorithms

(hyper-parameter settings, language biases, complexity)

• Core Data analysis phase

• Post-processing of acquired Concept definitions including validations

NPTEL

Video Course on Machine Learning

Professor Carl Gustaf Jansson, KTH

Thanks for your attention!

The next lecture 2.5 will be on the

topic:

Tutorial for Week 2