Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd Logo
Upload

Welcome to Scribd!

  • 49 views

    Convolutional Neural Networks: CMSC 733 Fall 2015 Angjoo Kanazawa

    Uploaded by

    Sanjeeb
    This document provides an overview of convolutional neural networks (CNNs). It begins by explaining that CNNs are a type of deep learning that learns feature representations directly from data in a supervised manner. It then discusses how CNNs build increasingly abstract features through multiple convolutional and pooling layers. CNNs incorporate shared weights and translation invariance properties suitable for images. The document reviews how CNNs have achieved state-of-the-art performance on tasks like ImageNet classification and object detection.

    Copyright:

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

    Convolutional Neural Networks: CMSC 733 Fall 2015 Angjoo Kanazawa

    Uploaded by

    Sanjeeb
    49 views55 pages
    This document provides an overview of convolutional neural networks (CNNs). It begins by explaining that CNNs are a type of deep learning that learns feature representations directly from data in a supervised manner. It then discusses how CNNs build increasingly abstract features through multiple convolutional and pooling layers. CNNs incorporate shared weights and translation invariance properties suitable for images. The document reviews how CNNs have achieved state-of-the-art performance on tasks like ImageNet classification and object detection.

    Original Title

    CNN

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    This document provides an overview of convolutional neural networks (CNNs). It begins by explaining that CNNs are a type of deep learning that learns feature representations directly from data in a supervised manner. It then discusses how CNNs build increasingly abstract features through multiple convolutional and pooling layers. CNNs incorporate shared weights and translation invariance properties suitable for images. The document reviews how CNNs have achieved state-of-the-art performance on tasks like ImageNet classification and object detection.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    49 views55 pages

    Convolutional Neural Networks: CMSC 733 Fall 2015 Angjoo Kanazawa

    Uploaded by

    Sanjeeb
    This document provides an overview of convolutional neural networks (CNNs). It begins by explaining that CNNs are a type of deep learning that learns feature representations directly from data in a supervised manner. It then discusses how CNNs build increasingly abstract features through multiple convolutional and pooling layers. CNNs incorporate shared weights and translation invariance properties suitable for images. The document reviews how CNNs have achieved state-of-the-art performance on tasks like ImageNet classification and object detection.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    You are on page 1of 55

    Convolutional Neural

    Networks
    CMSC 733 Fall 2015
    Angjoo Kanazawa
    Overview
    Goal: Understand what Convolutional Neural
    Networks (ConvNets) are & intuition behind it.
    1. Brief Motivation for Deep Learning
    2. What are ConvNets?
    3. ConvNets for Object Detection
    First of all what is Deep Learning?
    ● Composition of non-linear transformation of
    the data.
    ● Goal: Learn useful representations, aka
    features, directly from data.

    ● Many varieties, can be unsupervised or supervised.


    ● Today is about ConvNets, which is a supervised deep
    learning method.
    Recap: Supervised Learning

    Slide: M. Ranzato
    Supervised Learning: Examples

    Slide: M. Ranzato
    Supervised Deep Learning
    So deep learning is about learning
    feature representation in a
    compositional manner.
    But wait,
    why learn features?
    The Black Box in a
    Traditional Recognition Approach

    Feature Post-processing Classifier


    Preprocessing Extraction (Feature selection, (SVM,
    (HOG, SIFT, etc) MKL etc) boosting, etc)
    The Black Box in a
    Traditional Recognition Approach

    Hand
    Engin
    eered
    Feature Post-processing Classifier
    Preprocessing Extraction (Feature selection, (SVM,
    (HOG, SIFT, etc) MKL etc) boosting, etc)
    Feature Post-processing
    Preprocessing Extraction (Feature selection,
    (HOG, SIFT, etc) MKL etc)

    ● Most critical for accuracy


    ● Most time-consuming in development
    ● What is the best feature???
    ● What is next?? Keep on crafting better features?
    ⇒ Let’s learn feature representation directly
    from data.
    Learn features and classifier
    together
    ⇒ Learn an end-to-end recognition system.
    A non-linear map that takes raw pixels directly
    to labels.

    Slide: M. Ranzato
    Building a complicated function

    Each box is a simple nonlinear function


    Building a complicated function

    Slide: M. Ranzato
    Building a complicated function

    ● Composition is at the core of deep learning methods


    ● Each “simple function” will have parameters subject to
    learning Slide: M. Ranzato
    Intuition behind Deep Neural Nets

    Slide: M. Ranzato
    Intuition behind Deep Neural Nets

    Slide: M. Ranzato
    Intuition behind Deep Neural Nets

    Layer 1 Layer 2 Layer 3 Layer 4

    The final layer outputs a probability distribution of categories.


    Slide: M. Ranzato
    A simple single layer Neural Network
    Consists of a linear combination of input
    through a nonlinear function:

    W is the weight parameter to be learned.


    x is the output of the previous layer
    f is a simple nonlinear function. Popular choice is max(x,0),
    called ReLu (Rectified Linear Unit)
    1 layer: Graphical Representation
    f h is called a neuron,
    hidden unit or feature.

    f
    f
    Joint training architecture overview
    Neural Net Training

    Slide: M. Ranzato
    Neural Net Training

    Slide: M. Ranzato
    Neural Net Training

    Slide: M. Ranzato
    Neural Net Training

    Slide: M. Ranzato
    Neural Net Training

    Slide: M. Ranzato
    Neural Net Training

    Slide: M. Ranzato
    Neural Net Training

    Slide: M. Ranzato
    When the input data is an image..

    Slide: M. Ranzato
    When the input data is an image..
    Reduce connection to local regions
    Reuse the same kernel everywhere
    Because interesting
    features (edges) can
    happen at anywhere in
    the image.
    Convolutional Neural Nets
    Detail
    If the input has 3 channels (R,G,B),
    3 separate k by k filter is applied to
    each channel.

    Output of convolving 1 feature is


    called a feature map.

    This is just sliding window, ex. the output of


    one part filter of DPM is a feature map
    Using multiple filters
    Each filter detects features in
    the output of previous layer.
    So to capture different features,
    learn multiple filters.
    Example of filtering

    Slide: R.
    Fergus
    Building Translation Invariance
    Building Translation Invariance via
    Spatial Pooling

    Pooling also subsamples the image,


    allowing the next layer to look at larger
    spatial regions.
    Summary of a typical convolutional
    layer
    Doing all of this consists one
    layer.
    ○ Pooling and normalization is
    optional.
    Stack them up and train just like multi-
    layer neural nets.
    Final layer is usually fully connected
    neural net with output size == number of
    classes
    Compare this to SIFT
    Compare this to SIFT
    Revisiting the composition idea
    Every layer learns a feature detector by
    combining the output of the layer before.
    ⇒ More and more abstract features are learned
    as we stack layers.
    Keep this in mind and let’s look at what kind of
    things ConvNets learn.
    Architecture of Alex Krizhevsky et al.
    ● 8 layers total.
    ● Trained on Imagenet Dataset
    (1000 categories, 1.2M
    training images, 150k test
    images)
    ● 18.2% top-5 error
    ○ Winner of the ILSVRC-
    2012 challenge.

    Slide: R.
    Fergus
    Architecture of Alex Krizhevsky et al.
    First layer filters
    Showing 81 filters of
    11x11x3.
    Capture low-level
    features like oriented
    edges, blobs.

    Note these oriented edges are


    analogous to what SIFT uses to
    compute the gradients.
    Top 9 patches that activate each filter
    in layer 1

    Each 3x3 block shows


    the top 9 patches for
    one filter.
    Second Layer
    Second Layer

    Note how the previous


    low-level features are
    combined to detect a
    little more abstract
    features like textures.
    ConvNets as generic feature extractor
    ● A well-trained ConvNets is an excellent feature
    extractor.
    ● Chop the network at desired layer and use the output as
    a feature representation to train a SVM on some other
    vision dataset.

    ● Improve further by taking a pre-trained ConvNet and re-training it on a


    different dataset. Called fine-tuning
    One way to do detection with
    ConvNets
    Since ConvNets extract discriminative features,
    one can crop images at the object bounding
    box and train a good SVM on each category.
    ⇒ Extract regions that are likely to have
    objects, then apply ConvNet + SVM on each
    and use the confidence to do maximum
    suppression.
    R-CNN: Regions with CNN features

    Best performing method on PASCAL 2012


    Detection improving previous methods by 30%
    Slide: R. Girshick
    ConvNet Libraries
    • Cuda-convnet (Alex Krizhevsky, Google)
    • Caffe (Y. Jia, Berkeley, now Google)
    ● Pre-trained Krizhevsky model available.
    • Torch7 (Idiap, NYU, NEC)
    more around.

    You might also like