Page 74 - CITP Review

Combined, these two words take on a much richer technology meaning.

The evolution of AI begins with machine learning (ML). Machine learning is a computing term used to
describe the capacity of a computer to leverage a process or set of rules to be followed in calculations or
other problem-solving operations; it is able to modify itself without being preprogrammed. As more data
is compiled, more patterning is available. At the center of machine learning are underlying rules, also
known as algorithms, which are written by human programmers. Using the vending machine scenario
introduced in the IoT section, machine learning can occur by building an algorithm that captures data
pertaining to vending machine sales and weather. Additional layering can occur to further capture data
pertaining to local events occurring in the vicinity of the vending machine — pedestrian traffic, municipal
road construction projects, time of day, and so on. Now, foresights can be created that will affect the
vending machines’ strategic placement, sales, content selection, supply chain planning, and logistics to
allow for maximum asset optimization.

Deep learning (DL) focuses more specifically on a subset of ML tools and techniques, applying them to
25
solving essentially any problem that requires thought. Deep learning is unique for two reasons. First, it
programs itself, and second, learning occurs similar to the human brain. Let’s explore these
differentiators:

 Deep learning can be applied to any type of data — audio, video, speech, printed text — to produce
analyses and conclusions that appear to have been determined by humans. Over time, as it gains
greater experience and uses new data to learn and train itself, the system can improve its
performance, increasing the likelihood of correct classifications. Think of the Google image library as
a data set, or a medical school’s repository of heartbeat sounds. These sources are good examples
of inputs for deep learning. When a smaller subset of data is analyzed, conclusions can be drawn that
are not necessarily accurate. As more inputs are added, further classification occurs, hence, DL
improves the probability of a correct outcome next time.
 Deep learning occurs by way of artificial neural networks. The human brain is somewhat unorganized,
and yet, we are able to draw conclusions based upon complex neurological processes. An artificial
neural network responds to, for example, verbal nuance, context, and idiom, often lost in just written
words.
Deep learning is also susceptible to unfair prejudices, known as biases, because the underlying
algorithms are created by humans. These biases can creep in during one of a couple different stages:

 Problem statement. DL models are developed to identify an outcome, but some outcomes are
subjective and prone to bias. An example of potential bias in a problem statement could be around
employment “worthiness” based upon age, location, existing occupation, current income, or job
history; unfair prejudices can creep into the assumptions inherent in these variables.
 Data collection. Sometimes the collection is not representative of reality. This can easily occur when
collecting photos. A photo of a chocolate chip muffin can be modeled similar to a photo of a spotted
puppy, resulting in unfair biases. Another example of a data collection bias occurred with Amazon’s
DL model, which was trained to vet applicants by observing patterns in resumes submitted to the
company over a 10-year period; most came from men, a reflection of male dominance across the
tech industry.

25
See www.forbes.com/sites/bernardmarr/2016/12/08/what-is-the-difference-between-deep-learning-machine-
learning-and-ai/#5e66912526cf, accessed August 10, 2019.

© 2019 Association of International Certified Professional Accountants. All rights reserved. 2-28