Scribd
Upload
5 views8 pages

Deep Learning and Neural Networks_MBA08138

The document presents an overview of deep learning and convolutional neural networks (CNNs) for facial recognition, detailing the architecture and functionality of CNNs, including convolutional, pooling, and fully connected layers. It describes the training process using a dataset of 30,312 facial images, achieving an accuracy of 73% after 10 epochs, while noting potential accuracy loss due to data downsizing. The model's performance metrics and the time taken for training are also highlighted.

Uploaded by

Ashwin Shenoy
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
5 views8 pages

Deep Learning and Neural Networks_MBA08138

The document presents an overview of deep learning and convolutional neural networks (CNNs) for facial recognition, detailing the architecture and functionality of CNNs, including convolutional, pooling, and fully connected layers. It describes the training process using a dataset of 30,312 facial images, achieving an accuracy of 73% after 10 epochs, while noting potential accuracy loss due to data downsizing. The model's performance metrics and the time taken for training are also highlighted.

Uploaded by

Ashwin Shenoy
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 8

DEEP LEARNING AND

NEURAL NETWORKS
COMPUTE VISION: FACIAL RECOGNITION

A presentation
by Ashwin
Shenoy
Kranthi Kumar
Ravi Teja
Pranjul Gupta
Convolutional layers apply filters (kernels) to input
data to extract local patterns and features. These
filters slide over the input to learn spatial
hierarchies of features. CONVOLUTIONA
L NEURAL
Pooling layers down-sample the spatial dimensions
NETWORKS
of the input, reducing computational complexity
and making the network more robust to variations
These are complex deep
in scale and orientation. learning algorithms that are
Commonly used activation functions in CNNs inspired by the image
include Rectified Linear Unit (ReLU) to introduce recognition capability of a cat.
non-linearity and improve the network's ability to
learn complex patterns. CNNs are especially
Fully connected (dense) layers connect every powerful in handling image
data due to their ability to
neuron in one layer to every neuron in the next
automatically learn
layer, enabling the network to make final hierarchical features,
predictions based on learned features. recognizing patterns at
A flatten layer is often used to convert the output different levels of
of convolutional and pooling layers into a 1D vector abstraction.
before feeding it into fully connected layers.
COMPUTER VISION
Enhanced User
Experience
Entertainment
Delight

MY MACHINE
Securi
Personalized
services
Authentica

RECOGNIZES ME
ty tion
TO BUILD A FACIAL RECOGNITION
MODEL TO ACCURATELY IDENTIFY
AND CLASSIFY INDIVIDUALS BASED
ON FACIAL IMAGES
DATA AND DATA ENGINEERING

30,312 face images from the


DigiFace 1M database stored
on Google Drive
72 face images of 421
Identities
Images standardized to 244 x 244
pixels
Training Split 80% insample 20%
outsample
CONVOLUTIONAL NEURAL NETWORK
Full
Cony
ne
To d Layer cted
ecip s
ut her
Inp r patt g
Lay
e erns lobal
Optimiz
er
224
Output
Layer:
224 Softmax
RGB activation

Categorical Class- Metric for


entropy C on testing
v
For best fit (loss na l o l ut i o Max
calculation) L aye
rs o o l ing
P rs
Laye
SO HOW DOES THIS MODEL WORK?

Training
• Data fed in batches
Image Pre- •
using generators
Model trained for 10
processing epochs
• Calidation data is
• Standardization used to assess
of image generalization

dimensions.
• Rescaling pixel Model Architecture
values to the • Convolution layers
range [0, 1]. capture hierarchical
features
• Max pooling layers
reduce spatial
dimensions
• Full connected layers
decipher global
patterns
• The output layer
classifies the image
into categories
RESULTS

Loss of 1.45 Accuracy of 73%

Suspecting loss of
accuracy due to
downsizing data from
1.2 million to 30,312

Elapsed Time 1 day 14


hours 57 minutes and 6
seconds

You might also like