Transfer Learning
Transfer learning is a machine learning technique where a model developed for a particular task is
reused as the starting point for a model on a second task. This is especially useful in deep learning due to
the large amounts of data and computational resources required to train deep neural networks from
scratch. The key idea is that features learned by a model on a large dataset can be leveraged for a new,
often related, task, reducing training time and improving performance when data for the new task is
limited.
ImageNet is a large-scale visual database widely used in deep learning for training and benchmarking
models, particularly in image recognition tasks.
ImageNet is a dataset of over 14 million annotated images,
The images are categorized into thousands of classes, with each class representing an object category
(e.g., animals, objects, plants).
AlexNet: The first CNN to demonstrate the power of deep learning in large-scale visual recognition.
[Link] Net Architecture
The Alexnet has eight layers with learnable parameters.
The model consists of five layers with a combination of
max pooling followed by 3 fully connected layers and they
use Relu activation in each of these layers except the
output layer.
1.n-f+1 = feature maps if no padding
N+2p –f+1 if padding is applied
2.n-f/s+1….stride values above 1
3.n+2p-f/s +1
n==input image size
f=filter/kernel size
s=stride/jump
p=padding
Padding in CNN refers to the addition of extra pixels around the borders of the input
images or feature map.
Stride determines how many squares or pixels our filters skip when they move across the
image, from left to right and from top to bottom.
Why Alexnet is successful
[Link] Relu as activation function
[Link] Augmentation
[Link]
VGG 16 Architecture
VGG-16 Architecture
● The first two layers are convolutional layers with 3*3
filters, and first two layers use 64 filters that results in
224*224*64 volume as same convolutions are used. The
filters are always 3*3 with stride of 1
● After this, pooling layer was used with max-pool of 2*2
size and stride 2 which reduces height and width of a
volume from 224*224*64 to 112*112*64.
● This is followed by 2 more convolution layers with 128 filters.
This results in the new dimension of 112*112*128.
● After pooling layer is used, volume is reduced to 56*56*128.
● Two more convolution layers are added with 256 filters each
followed by down sampling layer that reduces the size to
28*28*256.
● Two more convolution layers are added with 256
filters each followed by down sampling layer that
reduces the size to 28*28*256.
● Two more stack each with 3 convolution layer is
separated by a max-pool layer.
● After the final pooling layer, 7*7*512 volume is
flattened into Fully Connected (FC) layer with
4096 channels and softmax output of 1000