Blog

1. Explain how you think system maps are useful in defining the workflow in AI projects.

Ans. 1. A system map helps us to find relationships between different elements of the problem scoped. It explains how the work should be done step by step in an AI project. It helps us to understand complex issues with multiple interdependent factors that affect each other. The relationship between the elements is represented by arrows where longer arrows represent a longer time for a change to happen.
For example, the concept of water cycle can be better defined as a system map.

2. What are the key differences between rule-based and learning-based AI?

Ans. 2. Difference between Rule-based AI and Learning-based AI modelling:
• Rule-based Approach: It refers to AI modelling where relationship or patterns in data are defined by the
developer. The machine follows the rules or instructions mentioned by the developer and performs its task
accordingly. To train our machine, we feed data into it and label each image accordingly. There is an explicit
y=f(x) established for a set of input x and output y. Testing of the machine can be done by comparing the
image with the trained data and according to the labels of the image.
• Learning-based Approach: It refers to AI modelling where the relationship or patterns in data are not clearly
defined by the developer. In this approach, random data is fed into the machine and it is left to the machine to
figure out patterns and trends on its own. Generally, this approach is followed when the data is unlabelled and
too random for a human to make sense out of it. Thus, the machine looks at the data, tries to extract similar
features out of it and clusters same data sets together. In the end, with the output, the machine tells us about
the trends which it observed in the training data.

3. At what stage of AI project cycle should we take care of ethics and biases?

Ans. 3. The design phase, when training data is used to develop machine learning models, is the stage where most of the underlying
biases seep in. Bias can be introduced by the selection or sampling of the training data itself. This may happen
unintentionally by excluding certain groups or data sets. So, when the resulting model gets applied to these groups, the
accuracy is inevitably lower than it is for groups that were included in the training data. Additionally, training data
usually requires labels used to “teach” the machine learning model during training. These labels often come from humans,
which introduce the risks of bias. It is especially important to assess variance in performance across sensitive factors.
Ans. 4. Classification: This is a rule-based AI modelling technique used to classify each item in a set of data into one predefined
set of classes or groups. In classification, the algorithm is able to determine which points in the data set belong to either side
of the classification function represented by the dotted line. Usually, the data set used for classification is labelled data
and is sorted as labelling is done. For example, if we want to train a model to identify if an image is of a mango or grapes,
we need to train it with multiple images of both mango and grapes along with their labels. The machine, then, classifies
images on the basis of the labels and predicts the correct label for test data.
Clustering: This is a machine learning model where the machine generates its own rules or algorithms unlike a rule-based
model. The data fed into such a model is usually unlabelled or random. The algorithms are generated on their own based
on data set and the machine needs to derive patterns or trends from the training data set to cluster the ones which follow the
same pattern. Unlike classification, the final output labels are not known in advance in Clustering.
Ans. 5. The key differences between Supervised learning and Unsupervised learning are:
(a) Supervised learning deals with labelled data where the output data patterns are known to the system. On the
other hand, Unsupervised learning works with unlabelled data and the machine has to come up with the underlying
relationship between input and output.
(b) Supervised learning is less complicated than Unsupervised learning and requires less processing.
(c) The outcome of Supervised learning is more accurate and reliable as compared to Unsupervised learning.
Ans. 6. Artificial neural networks were initially designed to function like human brain but our brain is far more efficient than
artificial neurons. The key differences between the two are as under:
(a) A human brain has around 86 billion neurons in it while an artificial neural network typically has 10–1000 neurons.
(b) A single neuron can process both input and output information in a human brain whereas for artificial
neurons, there are completely different layers of neurons for input and output.
(c) The neurons of human brain are connected asynchronously and have no fixed pattern to work whereas artificial
neural networks are having connected layers that can work in loops and compute one by one.
(d) The human brain consumes less power and might get tired of the information load but artificial neural networks
need a power supply work all the time and never get tired.
(e) Signals in a human brain move at a speed dependent on the nerve impulse but signals in artificial neural networks
carry on continuously with floating-point numbers of the synaptic weights and can control the speed of the function.
Ans. 7. A multilayer perceptron consists of an input layer, one or more hidden layers and an output layer. Every unit in a layer is
connected to all units of the next layer. The information is passed on to the input layer and an activation function is
used to get the output of that layer. The output of one layer is passed as an input to the next layer, which is propagated
further until the last layer. Perceptron model has several constraints:
(a) Output values of a perceptron can take on only one of two values (0 or 1).
(b) Perceptron can only classify linearly separable sets of vectors. If the vectors are not linearly separable, learning
will never reach a point where all vectors are classified properly.
(c) Perceptron does not even have the ability to learn a simple logical function like Boolean “XOR”.
Ans. 8. Although AI-enabled technology can act as a catalyst to achieve the SDGs, it may also trigger inequalities that may act as
inhibitors to some of the SDGs. For example,
(a) As AI can help identify areas of poverty using satellite images, it may also lead to additional qualification
requirements for any job, consequently increasing the inherent inequalities and acting as an inhibitor towards
the achievement of No Poverty SDG.
(b) Another area where AI is acting as an inhibitor is uneven distribution of AI technology. For instance, complex
AI-enhanced agricultural equipment may not be accessible to small farmers and, thus, may widen the gap between
them and large producers in more developed economies, consequently inhibiting the achievement of some targets
of SDG 2 on Zero Hunger.
(c) There is another shortcoming of AI in the context of SDG on gender inequality due to insufficient research on
algorithms that discriminate against women and other minorities.
Ans. 9. AI can improve education access and quality in the following ways:
(a) Using AI, teachers can offer personalized assistance to each student based on their learning needs in the form
of text, audio, video, etc.
(b) AI can provide intelligent and supportive learning environment through AI-powered apps which can help
students get real time response from their teachers.
(c) AI can produce smart content of superior quality which includes virtual content like video lectures, video
conferencing, etc.
Ans. 10. AI can make governments more transparent and smoother to operate in the following ways:
(a) AI-enabled apps and chatbots can help citizens connect and engage with governments and provide transparency
and better accountability of government.
(b) Access to information, contacting local leaders, tracing people, companies and assets across the globe that help
in smoother operations using AI.
Some of the countries already using AI in government:
(a) The Australian government’s Department of Human Services uses virtual assistants on parts of its website
to answer questions and encourage users to stay in the digital channel. This country also has a virtual assistant
named “Alex” on Australian Taxation Office’s website that was engaged in 1.5 million conversations and
resolved over 81% of enquiries at first contact.
(b) The U.S. Citizenship and Immigration Services uses a virtual assistant named “Emma” to answer questions
and assist its users and take them to the right page of the website.
(c) The Infocomm Development Authority of Singapore has a virtual assistant called “Ask Jamie” that can respond to
user queries by Natural Language Processing.