Scribd Logo
Upload

Welcome to Scribd!

  • 3 views

    ML Lec 14 LeNeT CNN Architecture

    Uploaded by

    8d24wc8sj2

    Copyright:

    © All Rights Reserved
    Download as PPTX, PDF, TXT or read online from Scribd

    ML Lec 14 LeNeT CNN Architecture

    Uploaded by

    8d24wc8sj2
    3 views14 pages

    Copyright

    © © All Rights Reserved

    Available Formats

    PPTX, PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Copyright:

    © All Rights Reserved
    Download as PPTX, PDF, TXT or read online from Scribd
    Download as pptx, pdf, or txt
    3 views14 pages

    ML Lec 14 LeNeT CNN Architecture

    Uploaded by

    8d24wc8sj2

    Copyright:

    © All Rights Reserved
    Download as PPTX, PDF, TXT or read online from Scribd
    Download as pptx, pdf, or txt
    of 14

    CNN Architecture

    • Some popular CNN models are:


    (i) LeNet
    (ii) AlexNet
    (iii) GoogleNet
    (iv) VGG
    (v) ResNet
    (vi) DenseNet
    (vii) ResNeXt
    (viii)EfficientNet
    LeNet
    • LeNet is one of the pioneering convolutional
    neural network (CNN) architectures, developed
    by Yann LeCun and his colleagues in the late
    1980s and early 1990s.
    • It was primarily designed for handwritten digit
    recognition, specifically for the MNIST dataset.
    • It was used by many banks for recognition of
    hand written numbers on cheques.
    • This architecture achieves an error rate as low as
    0.95% on test data, i.e., accuracy was more than
    99%.
    LeNet: Architecture
    • Input Layer:
    – Typically accepts 32x32 pixel grayscale images. MNIST images
    (28x28) are often zero-padded to this size.
    • Convolutional Layer 1 (C1):
    – Applies 6 convolutional filters of size 5x5, resulting in 6 feature
    maps of size 28x28.
    – Activation function: typically sigmoid or tanh.
    • Subsampling or Pooling Layer 1 (S2):
    – Applies average pooling (subsampling) with a 2x2 filter and a
    stride of 2, reducing the feature map size to 14x14.
    • Convolutional Layer 2 (C3):
    -- Applies 16 convolutional filters of size 5x5, producing
    16 feature maps of size 10x10.
    LeNet: Architecture
    • Subsampling Layer 2 (S4):
    -- Similar to S2, it uses average pooling to reduce the
    feature map size to 5x5.
    • Fully Connected Layer (C5):
    -- Flattens the output from S4 and connects to 120
    neurons.
    • Fully Connected Layer (F6):
    -- Connects to 84 neurons.
    • Output Layer:
    -- Outputs 10 neurons corresponding to the 10 digit
    classes (0-9).
    LeNeT5
    • There are many versions of LeNet architecture.
    • Following is the LeNet5 CNN
    LeNeT5

    • After first
    convolutional layer,
    the input image is
    convoluted to the
    size of 28x28.
    • There are 6 kernels,
    so the output
    feature map is of
    depth 6.
    LeNeT5
    • Second layer is the pooling layer where the size is
    reduced to half, i.e., 14 × 14, by 2 × 2 filter with stride 2.
    LeNet5

    • In the third layer, convolution occurs again, but this time with
    16 filters of 5x5 size, default pad=0 and stride=1
    • After this layer, the size of the input image is reduced to
    10x10x16.
    LeNet5

    • In fourth layer, the subsampling takes place, and the image


    size 10x10x16 is reduced to 5x5x16, by 2 × 2 filter with stride
    2.
    • The feature map of fourth layer is flattened into a vector of
    400 values.
    LeNeT5

    • Finally, the flattened vector passes through 3 fully connected


    layer for classification, where 1st FC layer consists of 120 nodes,
    2nd of 84 nodes, and the 3rd one which is the output layer
    consists of 10 nodes, one for each class (here, digit 0 to 9).
    Summary of LeNeT5 Architecture
    LeNet: Key Characteristics
    • Activation Functions: Originally used sigmoid
    or tanh, but modern adaptations often use
    ReLU.
    • Pooling: Utilizes average pooling, which was
    common at the time. Max pooling has since
    become more popular in later architectures.
    • Backpropagation: LeNet was one of the first
    networks to use backpropagation for training.
    LeNET
    • Impact and Applications
    (i) Legacy: LeNet laid the groundwork for more complex
    CNN architectures, influencing subsequent developments
    in computer vision.
    (ii) Use Cases: Beyond digit recognition, variations of LeNet
    have been adapted for various image classification tasks.

    • Summary:
    -- LeNet's architecture introduced key concepts in deep
    learning, such as convolutional layers, pooling, and end-to-
    end training, and remains a foundational model in the
    study of neural networks and computer vision.
    Thank You!

    You might also like