Video 1

What is perception?

Perception is the understanding, identification and interpretation of things by humans i.e. how far a
human understands or perceives a problem or an environment by using different senses like eye/vision,
speech, ear, etc. In deep learning, different methods are developed to mimic the capabilities of human
perception and to make the machines perceive like humans by using different forms of data like speech,
text, speech recognition, etc. in order to make it human like intelligent.

How deep learning is differ from machine learning?

Machine Learning and Deep learning both are subsets of Artificial Intelligence. Suppose we have a bigger
domain AI , then machine learning is a subset of AI and further Deep Learning is a subset of machine
learning. Machine Learning is an Artificial Intelligence approach. On the other hand Deep Learning is
Machine Learning approach

Describe some of human perceptions.

Some human perceptions for the understanding of the daily routine and his/her near abouts are by
using the following senses:

1. Vision (eye)
2. Speech
3. Ear (Speech recognition)
4. Text, etc.

Video 2

Why we said that in deep learning data is considered as oil?

Data is considered as oil because whenever we are talking to a company or an organization the most
important thing is data. Basically with the rapid increase in the algorithms of AI the importance of data
increases. If we have 5000 instances of data we can prepare a good classifier and if we have 10 million
instances with annotations then it becomes equal to human knowledge capabilities. These tell us the
importance of data that is why it is said that in deep learning data is oil.

What do we mean by the instances of data?

Instances means rows of data. When the data is collected in a file including many rows, then a single
row of that is said to be an instance of that data.

What is annotated data?

Annotated data means labeled data. Annotated data is he one which different labels are provided for
different data points and then the data is used for classification. For example, in an image different
bounding boxes are formed labelling the different objects then it is said to be an annotated/labelled
data.

Video 3

Write some application of Deep learning?

Some applications of deep learning are as follows:

1. Face Recognition
2. Speech recognition
3. Image recognition
4. Natural language processing
5. Recommendations (Like in Facebook and YouTube)
6. Betterment of user experience (by suggesting relevant data)
7. Virtual assistants

Is deep learning only solve big problems?

No, deep learning is suitable for both small and complex problems. For example, it can be used for face
detection of a single person or we can say small number of people which is a small problem and can also
be used by big companies like facebook for the betterment of user experience by using his/her data
which is a complex problem. So, I an be used for both type of problems and just needs a lot of data.

Video 4

Write steps of deep learning frame work.

Steps of deep learning frameworks are as follows:

1. Classification (All sub steps are working on single layer)

a. Logistic Classifier
b. Stochastic Optimization
c. Data and Parameters Tuning
2. Deep networks
a. DNs
b. Regularization
 3. Computer Vision
a. Convolutions
b. Neural Networks
4. Text Processing (or any time series data)
a. Recurrent Neural Networks
b. Embeddings
c. LSTM

What is difference in classification and regression?

The output variable in classification is categorical or discrete i.e. data is divided into different classes for
example in automation of feedback the classification will result in the form of like or dislike. It is about
predicting a label. On the other hand, in regression, the output variable is numerical or continuous and
an exact value is generated in the result of regression. It is about predicting a quantity.

How neurons work ? What are perceptron ( google it )

Neuron is a biological model used for the transmission of signals in human body for the working
of brain. It contain two main parts, dendrites and axons. In neurons, the dendrite receives
electrical signals from the axons of other neurons, processes it, and then move along the axons.
Perceptron, on the other hand, is an artificial neuron or we can say that a mathematical model
of a neuron in an artificial neural network. In perceptron, the electrical signals are represented
as numerical values.

Video 5

What is linear classification?

Linear Classification is assigning different labels to given set of inputs and it differentiates data between
different classes like binary classes. It uses the equation:

Wx + b = y

Where w = weights

x = Input vector

b = bias

y = output vector

w and b comes from the training data. Then the output is generated in the form of scores which is then
converted into probabilities using Softmax Function. And at the end the data is classified according to
the results..
What are hidden layers in deep learning? (google it)

In an artificial neural networks, hidden layer is the one present between the input and output layers,
where the artificial neuron takes the set of weighted inputs and process it through an activation function
to produce an output. These are the layers of mathematical functions and each layer is designed
to produce an output particular to an expected result.

What is bias and what is weight ? (google it ).

Bias means our prior knowledge for a specific event and is simply a constant value or a constant vector.
For example, if we have to predict the weather conditions (rainy or not), there could be prior knowledge
like if it is cloudy outside or a specific month in which it is more likely to rain then the prediction will be
positive. On the other hand, weights are the values which are assigned to the inputs when it is to be
transmitted between neurons. These are the coefficients of the equation. The training set is first
initialized with different set of weights and these are the optimized during the training period. We can
get both weights and bias from the training data by the number of instances of different inputs.