Deep Visual Learning

Jayanta Mukhopadhyay
Dept. of Computer Science and Engg.
 Deep learning
◼ Learning using a “deep” neural network
Classical ANN:
Only a few
Input Output hidden layers.

Input W(1),b(1) W(2),b(2) W(m),b(m) Output

Deep architecture: Many hidden layers.

 Deep learning for solving
vision problems
◼ Object recognition
◼ Object Localization
◼ Semantic Segmentation
◼ Video summarization
◼ Tracking objects
◼ … car

Bus
 Deep architecture: Why so
late in application?
◼ Concepts introduced in 80’s.
◼ Basic principles remain the same.
◼ Two major reasons.
◼ Availability of large scale annotated data.
◼ Penetration of internet and smart phones.
◼ Wide spread of social networking.
◼ Online shopping, etc.
◼ Advancement of computing power.
◼ High throughput GPU computing.
 Classical Image Classification

hand-crafted
feature
Classifier output Tiger?
extractor Algorithm
Cat?
• Edges • Bayesian Lion?
• SIFT/SURF key Point • LDA
• HOG Regional Features • SVM
• Motion Features, etc. • KNN
 Classification Challenges
◼ Very tedious and costly to develop hand-crafted
features to handle various challenges.

View Point variation Deformation Occlusion Intraclass Variation

Illumination Clutter Instances Scale

Highly dependent on one application, and not
transferable easily to other applications.
 Classification through deep
learning
Learning filters of feature extraction and also classifier.
Increasing level of abstraction

Low-level Mid-level High-level Trainable

features features features classifier output

*Feature visualization of convolutional net trained on ImageNet (Zeiler and Fergus, 2013)
 Supervised Learning

Data: (x, y) where x is data, y is label

Goal: Learn a function f to map x → y
Examples: Classification, Regression,
Object detection, Semantic
Segmentation, Image Captioning, etc.
 Supervised Learning
Data Driven Approach to learn the model in three
steps:
Step 1: Define Model

Predicted output
(image label) Input data Model weights
Model (Image pixels)
structure

Learn a parametric function f composed by weight

parameters w to classify Image x as class label y.
 Supervised Learning

Step 2: Collect data.

Training True
input output
 Supervised Learning

Step 3: Learn the model.

Total Loss = Data Loss + Regularization Loss
Predicted output

Loss function: Regularizer:

Learned Minimize average Measures Penalizes
weights loss over training “badness” of complex
set prediction models
 Loss

◼ A loss function tells how good our current

i

classifier is.
◼ Data loss: Model predictions should match
training data
◼ Softmax Loss (Multinomial Logistic Regression):
𝑠
𝑒 𝑦𝑖
𝐿𝑖 = −log( 𝑠𝑗 )
σ𝑗 𝑒

𝐿𝑖 = −log(𝑃 𝑌 = 𝑦𝑖 𝑋 = 𝑥𝑖 )
 Cross-entropy Loss

◼ Another form of softmax loss.

◼ 2-class entropy:
◼ -(y log(p)+(1-y) log(1-p)); p: Prob. (y=1|o)
◼ Multiclass:
Estimated Prob. of o
belonging to c

Binary indicator (1 if o
belongs to c, else 0). True Prob. of o
belonging to c

More general:
 Regularization Loss
i

◼Regularization Loss: Model should be

“simple”, so it works on test data as “W” is
not unique with just data loss.
◼ L2 Regularization (Weight Decay)
◼ L1 Regularization
◼ Elastic net (L1 + L2)
 How to find best weights ?

• Gradient descent
• Back propagation algorithm
 Gradient Descent
How to update weights?
Initialize w randomly
While true:
Compute gradient at current point

Move downhill a little bit:

Learning rate: How big

updating the weights at each each step should be
iteration
 Back Propagation

◼ Forward pass:
◼ Run graph “forward” to compute loss
◼ Backward pass:
◼ Run graph “backward” to compute gradients
with respect to loss
◼ Efficient to compute gradients for big,
complex models.
 Learning filters for feature
extraction
◼ Correlation with a mask or kernel

w1 w2 w3
w4 wc w5
w6 w7 w8

g ( x, y) = w1 f ( x − 1, y + 1) + w2 f ( x, y + 1) + w3 f ( x + 1, y + 1) + w4 f ( x − 1, y) +
wc f ( x, y) + w5 f ( x + 1, y) + w6 f ( x − 1, y − 1) + w7 f ( x, y + 1) + w8 f ( x + 1, y + 1)
 Convolution in neural
architecture
◼ Output of a neuron:
weighted sum of
inputs
◼ weights defined by a
kernel Σ
W
◼ Sparse connectivity
◼ Shared weights for
every node
◼ Sufficient to describe
the model by a kernel
 Convolution Layer
Filters always extend the full
32x32x3 image depth of the input volume
5x5x3 filter (kernel)

32
Convolve the filter
with the image
i.e. “slide over the
32 image spatially,
3 computing dot
products”.
Convolution Layer
32x32x3 image x
5x5x3 filter
32

1 number:
the result of taking a dot product
between the filter and a small
5x5x3 chunk of the image
32 (i.e. 5*5*3 = 75-dimensional dot
3 product + bias)
Locality!

Objects tend to have a local

spatial support.
Convolution Layer
32x32x3 image activation
5x5x3 filter map
32

28

convolve (slide) over all

spatial locations

32 28
3 Translation Invariance! 1
object appearance is
independent of location
Weight sharing!
Convolution Layer activation
32x32x3 image map
5x5x3 filter
32

28

convolve (slide) over all

spatial locations

32 28
3 1

Consider a second, green filter.

 Convolution Layer (CONV)
activation maps

32

28

Convolution Layer

32 28
6 # CONV
3 6

For example, if we had 6 5x5x3 filters, we’ll get 6 separate activation maps:

We stack these up to get a “new image” of size 28x28x6!

 Features of CONV
◼ Locality:
◼ objects tend to have a local spatial support
◼ Translation invariance:
◼ object appearance is independent of location
◼ Weight sharing
◼ units connected to different locations have the same

weights
◼ equivalently, each unit is applied to all locations

◼ weights of filters are invariant.

◼ Each unit output of filter is connected to a local

rectangular area in the input.
◼ – Receptive Field
 Non-Linear Layer

◼ Increase the nonlinearity of the entire

architecture without affecting the
receptive fields of the convolution layer.
◼ Commonly used in CNN is ReLU.
 Non Linearity: Activation Functions
- A few examples
Sigmoid ReLU (Rectified Linear Unit)

tanh(x) Leaky ReLU

 Convolutional Neural Networks (CNN)

32 28 24

….
CONV, CONV, CONV,
ReLU ReLU ReLU
e.g. 6 e.g. 10
5x5x3 5x5x6
32 filters 28 filters 24
3 6 10

A CNN is a sequence of convolution layers and

nonlinearities.
 Parameters involved in
convolution layer
◼ Input Volume size W1 x H1 x D1
◼ No. of filters K with size Fw x Fh x D1 convolved
with stride (Sw,Sh).
◼ Input zero padded by ( Pw, Ph ) on both sides.
◼ Output volume size W2 x H2 x D2?
◼ W2 = (W1 – Fw + 2Pw)/Sw + 1
◼ H2 = (H1 – Fh + 2Ph)/Sh + 1
◼ D2 = K
◼ Parameters ?
◼ (Fw * Fh * D1) * K weights + K biases
◼ d-th depth slice of output is the result of
convolution of d-th filter over the padded input
volume with a stride, then offset by d-th bias
 224x112x64
112x56x64

Pooling Layer (POOL)

◼ To progressively reduce the spatial size of the
representation.
◼ to reduce the amount of parameters and computation in the
network.
◼ to control overfitting.
◼ Pooling partitions the input image into a set of non-
overlapping rectangles.
◼ For each such sub-region, outputs an aggregated value of
the features in that region.
◼ Maximum value (Max pooling)
◼ Average value (Average pooling)
◼ Operates over each activation map independently
 Pooling Layer (POOL)
224x112x64
112x56x64
Pool

1 1 2 4
Maxpool with 2x2
5 6 7 8
filter and stride 2 6 8

3 2 1 0 3 4

1 2 3 4

Single depth slice

 Parameters involved in pooling
◼ Input Volume size W1 x H1 x D1
◼ Pool size Fw x Fh with stride (Sw,Sh).
◼ Output volume size W2 x H2 x D2?
◼ W2 = (W1 – Fw )/S + 1
◼ H2 = (H1 – Fh )/S + 1
◼ D2 = D1
◼ Parameters ?
◼ 0!
◼ Uncommon to use zero-padding in
Pooling layers.
 Fully Connected Layer (FC)
◼ Contains neurons that connect to the
entire input volume
◼ as in ordinary Neural Networks.
◼ Input volume to FC layer can also be
treated as Deep Features.
◼ If the FC layer is a classifier, the input to
FC can also be treated as feature vector
representation for the sample.
 LeNet: A typical example
◼ I/P→CONV→POOL→CONV→POOL→FC→FC→O/P
◼ Number of parameters: 60k

◼ Number of floating point operations per inference: 341k

◼ Sigmoid used for non-linearity.

Six 5*5 filters, 2*2 average Six 5*5 filters, 2*2 average
Stride 1 pooling, Stride 1 pooling,
Stride 2 Stride 2

*Y. Lecun et al, Proceedings of the IEEE, 1998

 Efficient computation with
smaller kernels
◼ Successive filtering with smaller sized kernels covers
equivalent receptive field area of a larger size.
◼ Stack of three 3x3 conv (stride 1) layers has same
effective field as one 7x7 conv layer
◼ Deeper with more non-linearity
◼ Usually deeper the model better the accuracy

◼ Till it overfits!

◼ Fewer parameters:
2 2 2 2
◼ 3*(3 C ) vs. (7 C ) for C channels

◼ Used in VGG network

◼ no. of layers: 13 / 16 / 19 with only 3x3 kernels
 Handling scale in feature
representation
◼ Concatenating multiscale feature descriptor
Output
Inception Module 28x28x(128+196+96+256)

Used in Google Net

57 Layers including 21
Conv and 1 FC Concat

1x1x128 3x3x196 5x5x96 3x3

Conv. Conv. Conv. Maxpool

Input
28x28x256
 Vanishing gradient problem
◼ Gradient becomes zero (vanishes) at deeper
layers!
◼ Learn residual mapping! Used in ResNet
No. of layers:
X 34/50/101/152
Layer #1
.
.
Identity
F(X) Layer #n

+ X
H(X)=F(X)+X
 Batch Normalization
◼ Normalizes input activation map to a layer
by considering its distribution over a batch
of training samples.
◼ Each dimension of the input feature map CONV
individually normalized
◼ To make Gaussian activation maps.
◼ Advantages BN
◼ Improves gradient flow through the network.
◼ Allows higher learning rates.
Non-
◼ Reduces the strong dependence on initialization. linearity
◼ Acts as a form of regularization.
◼ Usually inserted after FC / CONV layers,
and before non-linearity.
 Drop out
◼ Randomly dropping out nodes of network (at
hidden / visible layers) during training.
◼ Temporarily removing it from the network, along with
all its incoming and outgoing connections.
◼ To regulate overfitting, more effective for smaller
dataset.
◼ Simulates learning sparse representation in hidden
layers.
◼ Implementation
◼ Retain output of a node with a probability p.
◼ Typically within [0.5,1] at hidden layers and [0.8,1] in visible
layers.
 Learning weights with drop out
◼ Weights become larger due to drop out.
◼ Needs to be scaled at the end training.
◼ A simple heuristic.
◼ Outgoing weights of a unit retained with
probability p during training, multiplied by p at test
time.
◼ Scaling may be carried out during training
time at each weight update.
◼ No need to rescale weight for the test network.

Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, Ruslan Salakhutdinov, Dropout: A Simple Way to Prevent
Neural Networks from Overfitting ,JMLR, 15(Jun):1929−1958, 2014.
 Depthwise Separable
Convolutions
❑ Suppose, we have DF × DF × M input feature map,
DF × DF × N output feature map and Dk × Dk spatial
sized conventional convolution filters.

𝐷𝐹×𝐷𝐹×M 𝐷𝐹×𝐷𝐹×𝑁
❑What is the computational cost for such a
convolution operation?
—Dk · Dk · M · DF · DF · N
❑ What is the number of parameters?
—Dk · Dk · M · N Courtesy: Ankita Chatterjee
 Depthwise Separable
Convolutions
❑ Now, think of M filters which are D K × D K (not D K × D K × M )
and think each M of these filters are operated separately on
M channels of input of spatial size DF × DF
❑ Number of parameters? DK · DK · M

❑ What is the computational cost for such a convolution

operation? DK · DK · DF · DF · M

❑This operation is known as Depthwise Convolution operation.

Courtesy: Ankita Chatterjee

 Depthwise Separable
Convolutions
❑ What is the output shape now? DF × DF × M
❑ Where did the N (output channels) go?
- not there as depthwise convolution operates only on input channels.
❑ Now think about 1 × 1 traditional convolution on DF × DF × M
featuremap to get DF × DF × N output.
❑ What is the computation cost? 1 ·1 ·M ·DF ·DF ·N = DF ·DF ·M ·N

Used in MobileNet-V1
Alternate layers of
Conv and D-S Conv

What is the number of parameters? 1 ·1 ·M ·N

❑ This operation is called 1 × 1 pointwise convolution
Courtesy: Ankita Chatterjee
 MobileNet-V1

Image taken from: MobileNet Paper

Image taken from: MobileNet Paper Courtesy: Ankita Chatterjee

 Width and Resolution
Multiplier
❑ Used as parameters for scaling the model architectures.
❑Width multiplier α ∈ (0, 1] to thin a network uniformly at
each layer
- The number of input channels (from M): αM
- The number of output channels (from N): αN
- Computational cost? Dk ·Dk ·αM ·DF ·DF + DF ·DF ·αM ·αN
- Reduces roughly by α2
❑ Resolution multiplier ρ ∈ (0, 1] to reduce the image resolution
and the internal representation of every layer by by this
factor
❑ With width multiplier α and resolution multiplier ρ, the
computational cost? Dk ·Dk ·αM ·ρDF ·ρDF + ρDF ·ρDF ·αM ·αN
- Another reduction by ρ2 Courtesy: Ankita Chatterjee
 MobileNet-V1

Image taken from: MobileNet Paper

Image taken from: MobileNet Paper

Courtesy: Ankita Chatterjee

 EfficientNets: Compound
Model Scaling
❑ By balancing all dimensions of the network—width,
depth, and image resolution—against the available
resources to get the best overall performance.

❑ To perform a grid search to find the relationship between

different scaling dimensions of the baseline network under a
fixed resource constraint

❑ Apply those coefficients to scale up the baseline network to

the desired target model size or computational budget

Courtesy: Ankita Chatterjee

 Compound Model Scaling

Comparison of different scaling methods. Unlike conventional scaling methods

(b)-(d) that arbitrary scale a single dimension of the network, compound
scaling method uniformly scales up all dimensions in a principled way.
Image taken from: EfficientNet Paper
Courtesy: Ankita Chatterjee
 EfficientNet Architecture
❑ The effectiveness of model scaling also relies heavily on
the baseline network

❑ A new baseline network is developed by performing a

neural architecture search using the AutoML MNAS
framework, which optimizes both accuracy and efficiency
(FLOPS).

❑ The resulting architecture uses mobile inverted bottleneck

convolution (MBConv), similar to MobileNetV2 and
MnasNet, but is slightly larger

Courtesy: Ankita Chatterjee

 EfficientNet Architecture

The architecture for baseline network EfficientNet-B0 is simple and clean, making it easier to scale and generalize.
Image taken from: EfficientNet Paper

Courtesy: Ankita Chatterjee

 EfficientNet Architecture
Model Size vs. Accuracy

EfficientNet-B0:
the baseline network
developed by AutoML MNAS,

Efficient-B1 to B7:
obtained by scaling up the
baseline network.

In particular, EfficientNet-B7
achieves new state-of-the-art
84.4% top-1 / 97.1% top-5
accuracy, while being 8.4x
smaller than the best existing
CNN.
Courtesy: Ankita Chatterjee
Image taken from: EfficientNet Paper
 EfficientNet performance on
other baseline architectures

Scaling Up MobileNets and ResNet.

Image taken from: EfficientNet Paper Courtesy: Ankita Chatterjee
 EfficientNet performance vs.
other models

Image taken from: EfficientNet Paper Courtesy: Ankita Chatterjee

 EfficientNet performance on other
datasets using transfer learning

Image taken from: EfficientNet Paper Courtesy: Ankita Chatterjee

 Training steps:
• Preprocessing of training dataset.
◼ Normalize data.
◼ Decorrelate data (Diagonal Covariance
Matrix).
◼ Whitening of data (Identity Covariance
Matrix).
◼ Subtract Mean.
 Training steps:
• Data augmentation.
◼ Horizontal Flips
◼ Random Crops on scaled input
◼ Color jitter
◼ Distortions
◼ Transformations
• Weight initialization
• Train the network by update of the weight
parameters.
 Few Training Tips
◼ Start with small regularization and find learning rate that
makes the loss go down.
◼ Can overfit very small portion of the training data.
◼ Train first few epochs with few samples to initiate the
hyper-parameters.
◼ If big gap between training accuracy and validation
accuracy, then it is overfitting.
◼ Try increase regularization.

◼ If no gap, then may increase model capacity.

 Transfer Learning

◼ No need of a lot of a data to train a CNN.

◼ Pre-trained models can be initialized for
CNNs at the early stage of training.

*Donahue et al, “DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition”, ICML 2014
Razavian et al, “CNN Features Off-the-Shelf: An Astounding Baseline for Recognition”, CVPR Workshops 2014
Transfer Learning Lower learning rate when fine-tuning;
1/10 of original LR good starting Point.
1. Train
on Data 2. Train on 3.Bigger With bigger
Set #1 smaller data set dataset dataset, train
FC –3 FC –3 Reinitialize FC –3 more layers
this layer Train
FC-2 FC-2 FC-2
and train these
FC-1 FC-1 FC-1 layers

CNN 4 CNN 4 CNN 4

CNN 3 CNN 3 Freeze CNN 3

these
POOL1 POOL1 POOL1 Freeze
these
CNN 2 CNN 2 CNN 2

CNN 1 CNN 1 CNN 1

 Object Recognition and Localization

Predicts
(i) Coordinates of
bounding boxes
of objects.
(ii) Class probs.
Two stage processing: Localization
and recognition

*Girshick et al, “Rich feature hierarchies for accurate object detection and semantic segmentation”, CVPR 2014.
 Two stage processing: Object
Recognition and Localization

Region Feature Classification

Proposal Extraction

Pre-CNN Exhaustive Hand Crafted Linear

RCNN Region Proposal CNN Linear SVM

Fast RCNN Region Proposal Deep

Faster RCNN Deep

J. Redmon, et al, "You Only Look Once: Unified, Real-Time Object Detection," 2016 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 2016, pp. 779-788.

Single stage processing:

◼ YOLO (You Only Look Once)

1. Divide into SxS non-overlapping grid. End to end training

2. For each grid, predict B boxes containing for both predicting
objects (x,y,w,h) with confidence score. bounding boxes and
object classes.
3. Get softmax probs. of object classes using
a pre-trained network for the box with Final output:
SxSX(5B+C) tensor.
max. confidence with non-maximal
suppression.
 YOLO architecture and loss
function
◼ A series of convolution layers (coupled with max
pooling) followed by a few fully connected
layers.
◼ Typically: For image of size 448x448
◼ S=7, B=2, C (no. of classes): 20
◼ Loss function:
◼ Localization loss
◼ Confidence loss
◼ Classification loss
◼ All of them use MSE.
 Semantic Segmentation
◼Label each pixel in the image with a
category label.
◼Don’t differentiate instances, only care
about pixels.
Instance Level Semantic Segmentation
◼Even differentiate instances
 Semantic Segmentation
Building Blocks of CNNs:
◼Convolution

◼Down-Sampling

◼ MaxPool, AvgPool, Strided Convolution (S>1)

◼ Up-Sampling
◼ UnPooling, Upconvolution
 Up- Sampling: Max Unpooling
1 2 6 3
MaxPool 5 6
3 5 2 1
1 2 2 1 7 8
0 0 1 0
7 3 4 8
After few layers 0 1 0 0
in Network
0 0 0 0
0 0 b 0 1 0 0 1
0 a 0 0 a b
0 0 0 0
Max Unpool
Pooling Indices
c d
c 0 0 d

Noh et al, “Learning Deconvolution Network for Semantic Segmentation”, ICCV 2015
 Upsampled convolution

2x2 5 0 6 0
5 6 Upsample 0 0 0 0

7 8 7 0 8 0
0 0 0 0
2 1.1 2.4 .6
Input
1.2 1.3 1.4 .7 Filter .05 .1 .05
.28 1.5 3.2 .8 .1 .4 .1
.7 .75 .8 .4 .05 .1 .05
 Fully Convolutional Network
(FCN)
◼ No FC layer, only CNNs.
◼ Dense prediction
◼ Downsampled output upsampled followed by
1x1 convolution for providing softmax score at
every pixel.
◼ No. of channel at the output layer=No. of
classes.
◼ Cross-entropy loss function.
 Semantic Segmentation
◼ The upsampling of learned low resolution semantic
feature maps is done using upconvolutions which are
initialized w
◼ ith billinear interpolation filters.

Long, Shelhamer, and Darrell, “Fully Convolutional Networks for Semantic Segmentation”, CVPR 2015
 Fusion of prediction
◼ Combines prediction of higher layer with lower layer
(summing them).
◼ Performs at 3 stages.

Long, Shelhamer, and Darrell, “Fully Convolutional Networks for Semantic Segmentation”, CVPR 2015
 Segnet: Encoder-Decoder
Network
Encoder
◼Takes an input image and generates a high-dimensional

feature vector
◼Aggregate features at multiple levels

Decoder
◼Takes a high-dimensional feature vector and generates a

semantic segmentation mask

◼Decode features aggregated by encoder at multiple levels.

◼Semantically project the discriminative features (lower

resolution) learnt by the encoder onto the pixel space

(higher resolution) to get a dense classification.
 Semantic Segmentation
(Encoder-Decoder)

SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation, Vijay Badrinarayanan, et al. PAMI, 2017
 U-Net
◼ Inspiration from wavelet analysis and
synthesis of signals and images. x’(n)
g(n) g’(n)
x(n) xd1(n)
h(n) h’(n)
xa1(n) xd2(n)
g(n) g’(n)

h(n) h’(n)
xa2(n)
 U-Net architecture
◼ At the final layer a 1x1 convolution used to map each
feature vector to the desired number of classes.

Conv Concat

D/S U/S
Pool Conv
Encoder Decoder
U-Net: Convolutional Networks for Biomedical Image Segmentationm, Olaf Ronneberger, Philipp
Fischer, Thomas Brox, https://arxiv.org/abs/1505.04597
Results Ground truth
(Manual)
Yellow border
Color parts:
Segmented results
 Convolutional Encoder
Decoder
Self supervised
Typical architecture Learning

Auto encoder: Decoding the

Loss function: MSE same image.
 Encoder - Decoder:
Applications
◼ Denoising
◼ Segmentation
◼ Colorization
◼ Super-resolution
◼ …
Summary

◼ Deep architecture works in the same principle

of artificial neural network.
◼ A large number of hidden layers.
◼ A large number of weights.
◼ Convolution Neural Network (CNN)
◼ Learns filter weights.
◼ Sharing of weights.
◼ Two types of layers
◼ Convolutional and Pooling Layers
◼ Two stages
◼ Feature extraction
◼ Classification
 Summary
◼ Region proposal networks for
simultaneous object localization and
classification.
◼ YOLO – end to end single staged network.
◼ Fully Convolutional Networks
◼ Semantic Segmentation
◼ U-Net
◼ Encoder Decoder Network