Deep learning for librarians
| DOI | https://doi.org/10.1108/LHTN-09-2019-0067 |
| Pages | 16-22 |
| Date | 21 November 2019 |
| Published date | 21 November 2019 |
| Author | Donna Ellen Frederick |
| Subject Matter | Library & information science |
Deep learning for librarians
Donna Ellen Frederick
Over the past few years, the Data
Deluge Column has investigated various
ways in which new technologies make
use of the increasingly connected and
voluminous body of data available
through electronic networks. Librarians
have an established interest in
developments in areas such as linked data
and open science. However, there are
other areas of technological innovation,
which exist only on the periphery of
library and information science (LIS).
One of these areas is that of artificial
intelligence (AI). While the technologies
we use in libraries may either currently or
inthefuturemakeuseofAI,itisnotyet
part of the mainstream thinking in LIS
theory and practice. We see hints of AI,
for example, in the electronic resource
platforms and discovery services, which
have elements of a “recommender
service”. These features track our
searching and reading interests to make
predictions about other articles or books
that might also be of interest to us. In the
world of AI, recommenders are generally
seen as commonplace and may not even
requiretheuseof trueAIprocesses.In
this instalment of the Data Deluge
Column, we will take a deep dive into the
topic of AI and look at deep learning
(DL), how it works and what significance
it may have for the field of LIS.
To begin, the Merriam-Webster
Dictionary defines AI as “a branch of
computer science dealing with the
simulation of intelligent behavior in
computers”. This definition reveals that
AI is a very broad area of study and
innovation. A subdiscipline within AI is
machine learning (ML). The dictionary
definition for ML is “the process by
which a computer is able to improve its
own performance(as in analyzing image
files) by continuously incorporating new
data into an existing statistical model”.
While these definitions are relatively
easy to understand on the surface, they
leave the author of this column
wondering about the process to achieve
all of this. How, for example, can
machines that essentially function on a
binary principle simulate intelligence?
Intelligence implies the ability to reason
in a complex or abstract manner. This
binary environmentis one where all data
or information can be expressed and
processed byfirst converting it into zeros
and ones. Informationmust be simplified
in order to be processed this way. When
we think about “intelligence” and the
ability to “improve performance”, there
appears to be an implication of much
more complex “processing” of data than
what occurs in the traditional model of
information processing. Essentially,
computers or networks and algorithmsor
programs must become “intelligent” and
“learn” fromdata rather than just process
it. This is an intriguing proposition.How
can a machine, onewhich we understand
functions on a “garbage in, garbage out”
principle, learn how to improve its
performance simply by adding more data
to process? There needs to be a process
or model in which computers have the
capacity to learnin complex and abstract
ways. This is where the discipline of AI
and ML branchout into the area of DL.
Interestingly, there is no definition for
DL in the Merriam-Webster Dictionary or
at least not in the edition consulted by the
author of this column. Perhaps, this is an
indication that DL is not as mainstream in
popular culture as are AI and ML. Instead,
a technology encyclopedia needed to be
consulted to find a reasonably robust and
useful definition. Technopedia defines DL
as “a collection of algorithms used in
machine learning, used to model high-
level abstractions in data through the use
of model architectures, which are
composed of multiple nonlinear
transformations. It is part of a broad family
of methods used for machine learning that
are based on learning representations of
data.” While this definition is somewhat
informative, it does not add much to what
was discussed in the previous paragraph.
In fact, it seems to create an impression
that, within the realm of AI, DL is
somewhat diminutive. It is only a
“collection of algorithms” rather than a
fully developed approach for building AI
in real-world environments. In doing some
background learning to prepare for writing
this column, the author discovered that DL
is a complex and fascinating area of
computer science. It is one area of study
that brings together the shared interests of
many disciplines, including teaching,
cognitive psychology, information
science, neuroscience, engineering and
many others. While a thorough
explanation of what DL is and how it
works is both beyond the scope of this
column and the technical expertise of this
author, the basic theory and principles of
DL are relatively accessible and
interesting.
To begin, it is important to
recognize that while DL is a type of
ML, it has some critical differences.
With ML, it is necessary to program a
system how to identify things. With
DL, patterns in observed data are
processed through a type of electronic
neural network. This is essentially a
different process than setting out to
program a machine to do the
identification. A second difference is
that ML is essentially human-directed
while the goal of DL is to use little to
no human interaction in the “learning”
process. To achieve the learning, DL
requires a substantial amount of
carefully selected and presented data.
This is where the heavy work of
humans occurs in DL systems rather
than in activities associated with
programming.
DL is inspired by the functioning of
the human brain. This is an interesting
source of inspiration. The human brain
is a messy thing full of neurons,
synapses, neurotransmitters and all sorts
of electrical activity. It can be a fickle
and confounding thing at times! It is
also an amazing organ with an
incredible capacity to learn and be
creative. If not for the amazing powers
of the human brain, we would have no
computers, no art, no society and no
16 LIBRARY HITECH NEWS Number 1 2020, pp. 16-22, V
CEmerald Publishing Limited, 0741-9058, DOI 10.1108/LHTN-09-2019-0067
To continue reading
Request your trial