1. What is convolution Network? Explain about Convolution operation and 2.

2. List Examples of data types with different dimensionalities and number

Pooling. of Channels.
Convolutional networks (LeCun, 1989), also known as convolutional neural The data used with a convolutional network usually consists of several
networks or CNNs, are a specialized kind of neural network for processing channels, each channel being the observation of a different quantity at some
data that has a known, grid-like topology. Examples include time-series data. point in space or time.
Convolution is a mathematical operation that combines two functions to Single channel Multi-channel 1-D Audio waveform: The axis we convolve
produce a third function, expressing how the shape of one is modified by the over corresponds to time. We discretize time and measure the amplitude of
other. It is widely used in signal processing, computer vision, and machine the waveform once per time step. Skeleton animation data: Animations of 3-
learning, particularly in convolutional neural networks (CNNs). In the D computer-rendered characters are generated by altering the pose of a
continuous domain, convolution is defined as: s(t)=(x∗w)(t )= ∫−∞∞ “skeleton” over time. At each point in time, the pose of the character is
x(a)w(t−a)da Input Function (x): Represents the signal or data to be described by a specification of the angles of each of the joints in the
processed. In the example, x(t)x(t)x(t) is the position of a spaceship character’s skeleton. Each channel in the data we feed to the convolutional
measured by a noisy laser sensor. Kernel Function (w): A weighting function model represents the angle about one axis of one joint. 2-D Audio data that
that assigns more significance to recent measurements. The result s(t)s(t)s(t) has been preprocessed with a Fourier transform: We can transform the
is a less noisy estimate of the spaceship’s position, achieved by averaging audio waveform into a 2D tensor with different rows corresponding to
measurements weighted by w. Pooling is a process used in convolutional different frequencies and different columns corresponding to different points
neural networks (CNNs) to reduce the spatial dimensions of feature maps in time. Color image data:. The convolution kernel moves over both the
while retaining important information. Pooling helps to make the horizontal and vertical axes of the image, conferring translation equivariance
representation more compact, invariant to small transformations, and in both directions.3-D Volumetric data: A common source of this kind of
computationally efficient for subsequent layers. Pooling replaces the values data is medical imaging technology, such as CT scans. Color video data: One
of the feature map in a specific region (a pooling window) with a summary axis corresponds to time, one to the height of the video frame, and one to
statistic. For a 2×22 \times 22×2 pooling window:mpu:4,apu:2.5. the width of the video frame.

Input: [1234],Max Pooling Output: 4,Average Pooling Output: 2.5\text{Input:

} \begin{bmatrix} 1 & 2 \\ 3 & 4 \end{bmatrix}, \quad \text{Max Pooling
Output: } 4, \quad \text{Average Pooling Output: } 2.5Input: [1324
3. Describe
],Max Poolingabout : i) LeNet
Output: ii) AlexNet.
4,Average Pooling Output: 2.5 4. Discuss about RNN and how to compute the gradient in RNN?
LeNet LeNet is one of the earliest and most influential convolutional neural Recurrent Neural Networks (RNNs) are a type of neural network designed for
networks (CNNs), developed by Yann LeCun. Originally designed for sequential data, where the output at each time step depends not only on
handwritten digit recognition, LeNet became famous for its ability to classify the current input but also on the previous time steps. This is achieved by
digits from the MNIST dataset. The architecture consists of several maintaining a hidden state vector h(t)h(t)h(t) that captures the context of
convolutional layers followed by pooling layers, culminating in fully previous inputs. The fundamental task in training an RNN is to compute the
connected layers. LeNet's key components include the use of convolutional gradients for the network's parameters (such as weights WWW, UUU, and
layers for automatic feature extraction, subsampling (pooling) layers for VVV) through the process of backpropagation through time (BPTT).
dimensionality reduction, and a softmax classifier at the end for output. To compute the gradient in RNNs, we unroll the network over time, treating
Although relatively simple by modern standards, LeNet laid the foundation the time steps as a sequence of operations that can be represented as a
for many subsequent advancements in deep learning and computer vision. computational graph. The BPTT algorithm applies the standard
AlexNet AlexNet significantly improved performance on the ImageNet backpropagation technique recursively through this unrolled graph. Starting
dataset by using deeper networks, more sophisticated data augmentation from the output at the final time step τ\tauτ, the gradient of the loss with
techniques, and advanced GPU-based parallelization. The architecture respect to the output o(τ)o(\tau)o(τ) is calculated first. This gradient is then
consists of 8 layers: 5 convolutional layers followed by 3 fully connected propagated backward through the sequence, adjusting the hidden state
layers. It uses Rectified Linear Units (ReLU) for activation functions, which gradients h(t)h(t)h(t) by combining the gradients from both the next time
greatly improved training speed. Additionally, it introduced dropout step and the current output. The gradient of the loss with respect to the
regularization to prevent overfitting and employed local response output at time t, y(t) is: ∂L∂y(t)=y^(t)−y(t) Where y^(t)\hat{y}^{(t)}y^(t) is the
normalization (LRN) for better generalization. AlexNet's success marked the predicted output (often obtained via softmax) and y(t)y^{(t)}y(t) is the true
resurgence of deep learning, showing that CNNs could be highly effective for target.
large-scale image classification tasks.

.
5. Explain in detail about Recursive Neural Network?
Recursive Neural Networks (RNNs) are a class of neural networks that
extend the idea of recurrence (as seen in traditional RNNs) to tree-
structured data. Unlike recurrent neural networks that process data in a
linear sequence, recursive neural networks (RvNNs) operate on hierarchical
structures like trees. This architecture is particularly useful for processing
data that naturally forms a tree-like structure, such as sentences in natural
language, parse trees, and other hierarchical representations in both natural
language processing (NLP) and computer vision.
The computational graph for a recursive neural network is fundamentally
different from the chain-like structure of RNNs. In an RvNN, each node in the
graph represents a composition of two or more substructures, making it a
tree. This tree structure can either be predefined or learned from data. For
example, a parse tree for a sentence is one such structure that RvNNs can
operate on, where each non-leaf node combines two subtrees (child nodes),
and the leaves represent the atomic elements (words in the case of
sentences). Key Characteristics: Tree-Based Structure: RvNNs are designed
for tasks where the input has a hierarchical or tree-like structure, Efficient
Depth: One advantage of recursive neural networks over recurrent ones is
their ability to process sequences with fewer layers., Data-Dependent or
Predefined Tree Structures: The tree structure in recursive networks can be
either predefined or learned.
6. Explain Structured Outputs in CNN and Variants of basic Conv. Function. 7. List the three strategies for obtaining convolution kernels without
A structured output refers to an output that captures spatial relationships or supervised Training and Explain. ( Unsupervised Features)
dependencies within the data. Instead of providing a single output, CNNs Three Strategies for Obtaining Convolution Kernels Without Supervised
can emit a tensor where each element represents a prediction for a specific Training are: Random Initialization: Convolution kernels can be initialized
location or feature in the input. Structured outputs in Convolutional Neural with random values, rather than being trained through supervised learning.
Networks (CNNs) involve generating high-dimensional outputs, such as Despite appearing counterintuitive, random filters often work surprisingly
tensors, that maintain spatial relationships in the data. These outputs are well in convolutional networks. This method is computationally inexpensive
commonly used for tasks like image segmentation, where each pixel is and can be an effective way to experiment with different network
assigned a class label. For example, the CNN can output a tensor SSS, where architectures. After initializing the filters randomly, the final output layer is
Si,j,kS_{i,j,k}Si,j,k represents the probability that pixel (j,k)(j, k)(j,k) belongs to typically trained using supervised learning techniques, such as logistic
class iii. While pooling and striding often reduce spatial resolution, strategies regression or SVM, to classify the extracted features. Hand-Designed
like avoiding pooling, using unit-stride pooling, or emitting lower-resolution Kernels: In this strategy, convolution kernels are manually designed by the
grids can address this. CNNs can also refine predictions iteratively using researcher to perform specific tasks, such as detecting edges at particular
recurrent convolutional networks or enhance coherence through post- orientations or scales. This approach leverages domain knowledge to create
processing methods like segmentation into regions or graphical models. filters tailored to detect meaningful patterns in the data. For example,
Apps: Dense Prediction Tasks: Generating predictions that are spatially kernels could be designed to detect vertical or horizontal edges, or other
dense, like optical flow estimation or depth prediction. Contour Detection: basic visual features, before passing them to the network for further
Identifying object boundaries in images. Unlike standard convolution, which processing. Unsupervised Learning of Kernels In this method, convolution
shares the same kernel weights across the entire input, unshared kernels are learned in an unsupervised manner, without requiring labeled
convolution assigns separate kernels to different regions of the input. In data. One popular approach is to use techniques like k-means clustering to
tiled convolution, kernel weights are partially shared across different regions group similar patches from images, and the centroids of these clusters are
of the input. A tiling pattern is defined, and each tile has its own ker weights. then used as the convolution kernels. This approach allows the network to
learn filters that capture patterns in the data without the need for manual
design or supervised training. By training kernels unsupervised, the
convolutional network can first extract features from the data, and then a
classifier layer is trained separately using these features. This method
8. Explain about i ) Computer Vision ii ) Speech Recognition 9. Explain Bidirectional RNN ? Discuss about Deep Recurrent Networks.
reduces the computational burden during training by allowing feature
Computer Vision is a field of AI that enables machines to interpret and A Bidirectional Recurrent Neural Network (BiRNN) is an extension of the
extraction to be done once for the entire dataset, followed by a simpler,
understand visual information from the world, similar to human vision. It traditional Recurrent Neural Network (RNN) designed to capture
supervised classification step
involves tasks such as object recognition, image classification, face detection, dependencies from both past and future inputs in sequence-based tasks A
and image segmentation. Deep learning has significantly advanced computer BiRNN consists of two RNNs: A forward RNN that processes the input
vision by allowing models to learn complex features from raw data. sequence from time step 1 to time step TTT. A backward RNN that processes
Applications include identifying objects, annotating images with bounding the sequence in reverse, from time step TTT to time step 1. State
boxes, and labeling pixels in an image. Additionally, generative models are Representation: At each time step ttt, the forward RNN produces a hidden
used for image synthesis and restoration, such as repairing defects or state h(t)h^{(t)}h(t) based on the past inputs, while the backward RNN
removing objects from images. Computer vision is key to many AI produces a hidden state g(t)g^{(t)}g(t) based on future inputs. The final
applications, ranging from everyday tasks to innovative uses, like recognizing output at time ttt is a combination of these two hidden states, providing a
sound waves through vibrations in videos. Speech recognition is the process representation that depends on both the past and future context.
of converting spoken language into written text by mapping an acoustic Applications: Speech Recognition, Natural Language Processing (NLP). Deep
signal, typically represented as a sequence of input vectors, to a sequence of Recurrent Networks (Deep RNNs) refer to an extension of traditional
words or phonemes. The goal is to develop a function that computes the Recurrent Neural Networks (RNNs) that incorporate multiple layers of
most probable linguistic sequence given an acoustic input. Historically, hidden states between the input and the output. In Deep RNNs, the idea is
speech recognition systems relied on Hidden Markov Models (HMMs) and to introduce depth in the hidden layers of the network. Instead of having
Gaussian Mixture Models (GMMs) to model phoneme sequences and just one layer of hidden states, multiple layers of hidden states are used to
acoustic feature associations. However, in recent years, deep learning has capture more complex representations. This allows the network to learn
significantly advanced the field. more abstract features at higher levels of the hierarchy.

10. Explain LSTM with Block Diagrams.

Long Short-Term Memory (LSTM) is a type of recurrent neural network
(RNN) architecture that addresses the problem of learning long-term
dependencies in sequence data. The core contribution of LSTM is the
introduction of memory cells with self-loops, which allows the network to
retain information over long sequences. These memory cells are controlled
by several gating mechanisms, enabling the model to selectively forget,
update, and output information. LSTMs have shown significant success in
tasks like speech recognition, handwriting recognition, language modeling,
and machine translation. The LSTM network is composed of LSTM cells that
operate in a recursive manner, like standard RNNs, but with additional gates
and self-loops. These gates control the flow of information, allowing the
network to maintain a memory of previous states while adapting to new
inputs. The primary components are: Forget Gate: Determines what
information from the previous memory cell should be forgotten. Input Gate:
Decides what new information should be stored in the memory cell. Output
Gate: Controls the information sent to the next layer of the network.
Applications: Speech Recognition: LSTMs are used in systems that require
modeling of speech data, where long-term temporal dependencies are
critical, Machine Translation: LSTMs have been used to create powerful
sequence-to-sequence models that excel at translating sentences from one
language to another.