Image classification with Deep Neural Networks

Image Classification with
Deep Neural Networks
Yogendra Tamang
Sabin Devkota
Presented By:
February 6, 2016
ImageNet Classification with Deep Convolutional Neural Networks
A. Krizhevsky, I. Sutskever, G. Hinton
#pwlnepal
PWL Kathmandu
Papers We Love
Kathmandu

Outline
Background
Image Classification
Activation Functions
Convolution Neural Networks
Other Issues
Techniques used in other papers
And So on…

Background
Learning
Supervised
Unsupervised
AI Tasks
Classification and Regression
Clustering
Machine Learning Problem
Supervised
RegressionClassfication
Unsupervised
Clustering

Background
Classification
 Classifies data into one of discrete classes
 Eg. Classifying digits
 Cost Function for Classification Task may be Logistic Regression or Log-likelihood
Regression
 Predicts continuous real valued output
 Eg. Stock Price Predictions
 Cost function for regression type problem are MSE(Mean Squared Error)

Introduction
• Input Image is array of
number for computer
• Assign a label to input image
from set of categories
• One of core problems in
computer vision

Neural Networks and Multilayer Perceptrons

Neural Networks and Multi Layer Perceptrons
“activation” of unit in layer
matrix of weights controlling
function mapping from layer to
layer
They use Activation Functions.

ReLU Activation
non-saturating non-linearity
While training with Gradient
Descent these are faster than
saturating non-linearities like
sigmoid and tanh functions.

Convolutional Neural Networks
• Replicated feature approach
• Local Connectivity:
Receptive fields(RFs)
• Shared Weights:
Apply filter to image.
• Pooling:
subsampling layers

Convolutional Neural Networks
One or more convolutional layer
Followed by one or more fully connected layer
Resulting in easy to train networks with many fewer
parameters.

Learning a Classifier
Gradient Descent Algorithm
 Calculate Cost Function or Lost Function J(s)
 Calculate Gradient 𝜕𝐽(w)/𝜕w
 Update Weights
Stochastic Gradient Descent: Updates Adjust after example.
Mini-batch SGD: Updates after batches.
We do not use Squared Error measure for training instead we use Softmax
Function for output
Right cost function is Negative Log Likelihood.

Learning a Classifier- Negative Log
likelihood
𝑁𝐿𝐿 𝜃, 𝒟 = −
𝑖=0
|𝒟|
log 𝑃(𝑌 = 𝑦(𝑖)|𝑥 𝑖 , 𝜃)
Where 𝒟 is Dataset
𝜃 is weight parameter
(𝑥 𝑖
, 𝑦 𝑖
) is ith training data. Y is target data.

Overfitting
When the number of training
examples are small and the
architecture is deep, the network
performs well on training data but
worse on test data. i.e. it overfits.

Overfitting Mitigation
Data Augmentation :
Artificially creating more data samples from existing data through various
transformation of images (i.e. rotation, reflection, skewing etc.) and/or dividing
images into small patches and averaging all their predictions.
Applying PCA to the training examples to find out principal components which
correspond to intensity and color of the illumination. Creating artificial data by
adding randomly scaled eigen vectors to the training examples.
𝐼𝑥𝑦 𝑅, 𝐼𝑥𝑦 𝐺, 𝐼𝑥𝑦 𝐵 𝑇 = 𝒑𝟏, 𝒑𝟐, 𝒑𝟑 [𝛼1 𝜆1, 𝛼2 𝜆2, 𝛼3 𝜆3] 𝑇
Dropout:

Dropout
Technique to reduce overfitting
Dropout prevents complex co-adaptation on the
training data.
Randomly omit each hidden unit with probability
0.5 and
Its like randomly sampling from 2^H
Architectures, H is number of units in a hidden
layer
Efficient way to average many larger neural nets

ImageNet classification with Deep CNN
Improvement increases with larger datasets
Need model with large learning capacity
CNN’s capacity can be controlled with depth and breadth
Best results in ILSVRC-2010, 2012

CNN Architecture Design
5 Convolution Layers
3 Full Connection Layers
Last output is 1000 way

Training on Multiple GPUs
Current GPU more suited to cross-GPU Parallelization
Putting half of neurons on each GPUs and allowing them to communicate only in
certain layers
Choosing the connectivity is done by cross validation

Local Response Normalization
Activity of neuron i at position (x, y) is normalized to simulate lateral inhibition
inspired by real neurons.

Datasets
ILSVRC started as part of Pascal Visual Object challenge on 2010
1.2 Million training images, 50K validation Images and 150K testing images.
ILSVRC uses 1000 Images for each of 1000 categories.
Two error measures top-1 error, top-5 error.
Top-5 error fraction of test images for which correct label is not among the best 5
results predicted from the model.

Visualization of Learned feature

Tools used for overfitting
Data Augmentation
Dropout

References[1] A. Krizhevsky, I. Sutskever and G. E. Hinton, "ImageNet Classification with Deep Convolutional Neural
Networks," 2012.
[2] S. Tara, Brian Kingsbury, A.-r. Mohamed and B. Ramabhadran, "Learning Filter Banks within a Deep Neural
Network Framework," in IEEE, 2013.
[3] A. Graves, A.-r. Mohamed and G. Hinton, "Speech Recognition with Deep Recurrent Neural Networks,"
University of Toronto.
[4] A. Graves, "Generating Sequences with Recurrent Neural Networks," arXiv, 2014.
[5] Q. V. Oriol Vinyals, "A Neural Conversational Model," arXiv, 2015.
[6] R. Grishick, J. Donahue, T. Darrel and J. Mallik, "Rich Features Hierarchies for accurate object detection and
semantic segmentation.," UC Berkeley.
[7] A. Karpathy, "CS231n Convolutional Neural Networks for Visual Recognition," Stanford University, [Online].
Available: http://cs231n.github.io/convolutional-networks/.
[8] I. Sutskever, "Training Recurrent Neural Networks," University of Toronto, 2013.
[9] "Convolutional Neural Networks (LeNet)," [Online]. Available: http://deeplearning.net/tutorial/lenet.html.
[10] C. Eugenio, A. Dundar, J. Jin and J. Bates, "An Analysis of the Connections Between Layers of Deep Neural
Networks," arXiv, 2013.

References
[11] M. D. Zeiler and F. Rob, "Visualizing and Understanding Convolutional Networks," arXiv, 2013.
[12] G. Hinton, N. Srivastava, A. Karpathy, I. Sutskever and R. Salakhutdinov, Improving Neural Networks
by preventing co-adaptation of feature detectors, Totonto: arXiv, 2012.
[13] L. Fie-Fie. and A. Karpathy, "Deep Visual Alignment for Generating Image Descriptions,"
Standford University, 2014.
[14] O. Vinyals, A. Toshev., S. Bengio and D. Erthan, "Show and Tell: A Neural Image Caption
Generator.," Google Inc., 2014.
[15] J. M. G. H. IIya Sutskever, "Generating Text with Recurrent Neural Networks," in 28th International
Conference on Machine Learning, Bellevue, 2011.
[16] "Theano," [Online]. Available: http://deeplearning.net/software/theano/index.html. [Accessed 27
10 2015].
[17] "What is GPU Computing ?," NVIDIA, [Online]. Available: http://www.nvidia.com/object/what-is-
gpu-computing.html. [Accessed 27 12 2015].
[18] "GeForce 820M|Specifications," NVIDIA, [Online]. Available:
http://www.geforce.com/hardware/notebook-gpus/geforce-820m/specifications. [Accessed 28
10 2015].

Image classification with Deep Neural Networks

In this document