What is computer vision?

• The search of the fundamental visual features, and

the two fundamentals applications of reconstruction
and recognition
– Features, maps a 2D image into a vector or a point x
– Recognition
– Reconstruction
 2012
• The current mania and euphoria of the AI revolution
– 2012, annual gathering, an improvement of 10% (from 75 to 85)
• Computer vision researchers use machine learning techniques to recognize
objects in large amount of images
– Go back to 1998 (1.5 decade!)
• Textual (hand-written and printed) is actually visual!
– And why wait for so long?
• A silent Hardware revolution: GPU
• Sadly driven by video gaming 
– Nvidia (GPU maker) is now in the driving seat of this AI revolution!
• 2016, AlphaGo beats professionals
• A narrow AI program
• Re-shaping AI, Computer Science, digital revolution …
 Visual matching, and recognition for
understanding
• Finding the visually similar things in different images
--- Visual similarities

• Visual matching, find the ‘same’ thing under different

viewpoints, better defined, no semantics per se.
• Visual recognition, find the pre-trained ‘labels’,
semantics
– We define ‘labels’, then ‘learn’ from labeled data, finally
classify ‘labels’
 The state-of-the-art of visual
classification and recognition

• Any thing you can clearly define and label

• Then show a few thousands examples (labeled data)
of this thing to the computer
• A computer recognizes a new image, not seen
before, now as good as humans, even better!
• This is done by deep neural networks.
References
 In the notes.pptx, some advanced topics
 In talks, some vision history

 CNN for Visual Recognition, Stanford http://cs231n.github.io/

 Deep Learning Tutorial, LeNet, Montreal
http://www.deeplearning.net/tutorial/mlp.html

 Deep Learning, Goodfellow, Bengio, and Courville

 Pattern Recognition and Machine Learning, Bishop

 Sparse and Redundant Representations, Elad
 Pattern Recognition and Neural Networks, Ripley
 Pattern Classification, Duda, Hart, different editions

 A wavelet tour of signal processing, a sparse way, Mallat

 Introduction to applied mathematics, Strang

Some figures and texts in the slides are cut/paste from these references.
A few basics
 Vector spaces and geometries
Rn

E n  A n with dot prod.  P  R  pts at inf.
n n
 A n

Naturally everything starts from the known vector space

• add two vectors

• multiply any vector by any scalar
• zero vector – origin
• finite basis

7
• Vector space to affine: isomorph, one-to-one
Pts, lines, parallelism

• vector to Euclidean as an enrichment: scalar prod.

Angle, distances, circles

• affine to projective as an extension: add ideal elements

Pts at infinity

8
Numerical computation and Linear
algebra

• Gaussian elimination
• LU decomposition
• Choleski (sym positive) LL^T
• orthogonal decomposition
• QR (Gram-Schmidt)
• SVD (the high(est)light of linear algebra!)
• row space: first Vs
A m n  U  V T • null space: last Vs
• col space: first Us
• null space of the trans : last Us 9
 Applications

• solving equations A x = b, decomposition, SVD

• less, constraints, sparse solutions
• more, least squares solution, ||Ax – b||^2 is eigendecomposition or S
• solving linear equation, also iteratively nonlinear one f(x), or optimi
• PCA is SVD

10
 High dimensionality discussion --- curse

• [0,1]^d
•…

11
Visual recognition
 Image classification and object
recognition
• Viewing an image as a one dimensional vector of features x, thanks to features! (or simply pixels if
you would like),
• Image classification and recognition becomes a machine learning application f(x).

• What are good features?

• Unsupervised, given examples x_i of a random variable x, implicitly or explicitly we learn the
distributions p(x), or some properties of p(x)
– PCA Principal Components Analysis
– K-means clustering

• Given a set of training images and labels, we predict labels of a test image.

• Supervised, given examples x_i and associated values or labels y_i, we learn to predict a y from a x,
usually by estimating p(y|x)
– KNN K-Nearest Neighbors
• Distance metric and the (hyper) parameter K
• Slow by curse of dimension
– From linear classifiers to nonlinear neural networks
 Unsupervised

• K-means clustering
– Partition n data points into k clusters
– NP hard
– heuristics
• PCA Principal Components Analysis
– Learns an orthogonal, linear transformation of the data
 Supervised

• Given a set of training images and labels, we predict

labels of a test image.
• Supervised learning algorithms
– KNN K-Nearest Neighbors
• Distance metric and the (hyper) parameter K
• Slow by curse of dimension
– From linear classifiers to nonlinear neural networks
Fundamental linear classifiers
Fundamental binary linear classifiers

 Binary linear classifier

 The classification surface is a hyper-plane
 Decision function could be a nonlinear thresholding d(x_i,wx+b),
 Nonlinear distance function, or probability-like sigmoid

 Geometry, 3d, and n-d

 Linear algebra, linear space
 Linear classifiers

 The linear regression is the simplest case where we have a close

form solution
 A neuron is a linear classifier
• A single neuron is a linear classifier, where the decision function is
the activation function
• w x + b, a linear classifier, a neuron,

– It’s a dot product of two vectors, scalar product

– A template matching, a correlation, the template w and the input vector x
(or a matched filter)
– Also an algebraic distance, not the geometric one which is non-linear
(therefore the solution is usually nonlinear!)
• The dot product is the metric distance of two points, one data, the other
representative
• its angular distance, x . w = ||x|| || w|| cos theta, true distance d = ||x-w||^2 =
cosine law = ||x||^2 + ||w||^2- 2 ||x|| ||w|| x . x = x .x + w.w – 2 x.x w.w x . w
– The ‘linear’ is that the decision surface is linear, a hyper-plane. The decision
function is not linear at all
A biological neuron and its mathematical model.
 A very simple old example

• The chroma-key segmentation, we want to remove

the blue or green background in images
• Given an image RGB, alpha R + betta G + gamma B >
threshold or 0.0 R + 0.0 G + 1.0 B > 100
 A linear classifier is not that
straightforward
• Two things: Inference and Learning

• Only inference ‘scoring’ function is linear

– We have only one ‘linear’ classifer, but we do have different ways to define the loss to make the
learning feasible. The learning is a (nonlinear) optimization problem in which we define a loss
function.
• Learning is almost never linear!
– How to compute or to learn this hyperplane, and how to assess which one is better? To define an
objective ‘loss’ function

• The classification is ‘discrete’, binary 0, 1,

• No ‘analytical’ forms of the loss functions, except the true linear regression with y-y_i
• It should be computationally feasible
– The logistic regression (though it is a classification, not a regression), converts the output of the
linear function into a probability
– SVM
 Activation (nonlinearity) functions
• ReLU, max(0,x)
• Sigmoid logistic function ), normalized to between 0
and 1, is naturallly probability-like between 0 and 1,
– so naturally, sigmoid for two,
– (and softmax for N, )
– Activation function and Nonlinearity function, not
necessarily logistic sigmoid between 0 and 1, others like
tanh (centered), relu, …
• Tanh, 2 s(2x) – 1, centered between -1 and 1, better,
VisGraph, HKUST
 The rectified linear unit, ReLU
• The smoother non-linearities used to be favoured in
the past.
• Sigmoid, kills gradients, not used any more
• At present, ReLU is the most popular.
• Easy to implement max(0,x)
• It learns faster with many layers, accelerates the
learning by a factor of 6 (in Alex)

• Not smooth at x=0, subderivative is a set for a convex

function, left and right derivatives, set to zero for
simplicity and sparsity
From two to N classes

• Classification f : R^n  (1,2,…,n), while Regression f:

R^n  R
• Multi-class, output a vector function 𝒙′=𝑔(𝑾 𝒙 + 𝒃),
where g is the rectified linear.
• W x + b, each row is a linear classifier, a neuron
 The two common linear classifiers, with
different loss functions
• SVM, uncalibrated score
– A hinge loss, the max-margin loss, A loss for j is different from y
– Computationally more feasible, leads to convex optimization
• Softmax f(*), the normalized exponentials, 𝒚=𝑓(𝒙′ )=𝑓(𝑔(𝑾 𝒙+𝒃))
– multi-class logistic regression
– The scores are the unnormalized log probabilities

– the negative log-likelihood loss, then gradient descent

– (1,-2,0) -> exp, (2.71,0.14,1) -> ln, (0.7,0.04,0.26)
– ½(1,-2,0) = (0.5,-1,0) -> exp, (1.65,0.37,1) ->ln, (0.55,0.12,0.33), more
uniform with increasing regularization
• They are usually comparable
From linear to non-linear classifiers,
Multi Layer Perceptrons
• Go higher and linear!

• find a map or a transform 𝒙↦𝜙(𝒙) to make them linear, but

in higher dimensions
– A complete basis of polynomials  too many parameters for the
limited training data
– Kernel methods, support vector machine, …

• Learn the nonlinearity at the same time as the linear

classifiers  multilayer neural networks
 Multi Layer Perceptrons
• The first layer
• The second layer

• One-layer, g^1(x), linear classifiers

• Two-layers, one hidden layer --- universal nonlinear approximator
• The depth is the number of the layers, n+1, f g^n … g^1
• The dimension of the output vector h is the width of the layer

• A N-layer neural network does not count the input layer x

• But it does count the output layer f(*). It represents the class scores
vector, it does not have an activation function, or the identify
activation function.
A 2-layer Neural Network, one hidden layer of
4 neurons (or units), and one output layer with
2 neurons, and three inputs.

The network has 4 + 2 = 6 neurons (not

counting the inputs), [3 x 4] + [4 x 2] = 20
weights and 4 + 2 = 6 biases, for a total of 26
learnable parameters.
A 3-layer neural network with three inputs, two
hidden layers of 4 neurons each and one output
layer. Notice that in both cases there are
connections (synapses) between neurons across
layers, but not within a layer.

The network has 4 + 4 + 1 = 9 neurons, [3 x 4]

+ [4 x 4] + [4 x 1] = 12 + 16 + 4 = 32 weights
and 4 + 4 + 1 = 9 biases, for a total of 41
learnable parameters.
 Universal approximator

• Given a function, there exisits a feedforward network

that approximates the function
• A single layer is sufficient to represent any function,
• but it may be infeasibly large,
• and it may fail to learn and generalize correctly
• Using deeper models can reduce the number of units
required to represent the function, and can reduce
the amount of generalization error
Functional summary

 The input , the output y

 The hidden features
 The hidden layers is a nonlinear transformation
 provides a set of features describing
 Or provides a new representation for
 The nonlinear is compositional
 Each layer ), g(z)=max(0,z).
 The output
 MLP is not new, but deep feedforward
neural network is modern!

• Feedforward, from x to y, no feedback

• Networks, modeled as a directed acyclic grpah for
the composition of functions connected in a chain,
f(x) = f^(i)( … f^3(f^2(f^1(x))))
• The depth of the network is N
• ‘deep learning’ as N is increasing.
Forward inference f(x) and backward
learning \nabla f(x)
• A (parameterized) score function f(x,w) mapping the
data to class score, forward inference, modeling
• A loss function (objective) L measuring the quality of
a particular set of parameters based on the ground
truth labels
• Optimization, minimize the loss over the parameters
with a regularization, backward learning
The dataset of pairs of (x,y) is given and fixed. The weights start out as
random numbers and can change. During the forward pass the score
function computes class scores, stored in vector f. The loss function
contains two components: The data loss computes the compatibility
between the scores f and the labels y. The regularization loss is only a
function of the weights. During Gradient Descent, we compute the
gradient on the weights (and optionally on data if we wish) and use
them to perform a parameter update during Gradient Descent.
Cost or loss functions
• Usually the parametric model f(x,theta) defines a distribution p(y | x; theta) and we
use the maximum likelihood, that is the cross-entropy between the training data y
and the model’s predictions f(x,theta) as the loss or the cost function.
– The cross-entropy between a ‘true’ distribution p and an estimated distribution q is H(p,q) =
- \sum_x p(x) log q(x)
• The cost J(theta) = - E log p (y|x), if p is normal, we have the mean squared error
cost J = ½ E ||y-f(x; theta)||^2+const

• The cost can also be viewed as a functional, mapping functions to real numbers, we
are learning functions f parameterized by theta. By calculus of variations, f(x) = E_*
[y]

• The SVM loss is carefully designed and special, hinge loss, max margin loss,
• The softmax is the cross-entropy between the estimated class probabilities
e^y_i / \sum e and the true class labels, also the negative log likelihood loss L_i = -
log (e^y_i / \sum e)
MLE, likelihood and probabity discussions
KL, information theory, ….
Gradient-based learning

• Define the loss function

• Gradient-based optimization with chain rules
• z=f(y)=f(g(x)), dz/dx=dz/dy dy/dx
• In vectors and , the gradient , where J is the Jocobian matrix of
g
• In tensors, back-propagation

• Analytical gradients are simple

– d max(0,x)/d x = 1 or 0, d sigmoid(x)/d x = (1 – sig) sig

• Use the centered difference f(x+h)-f(x-h)/2h, error order of

O(h^2)
Stochastic gradient descent

• min loss(f(x;theta)), function of parameters theta,

not x
• min f(x)
• Gradient descent or the method of steepest descent
x’ = x - \epsilon \nabla_x f(x)
• Gradient descent, follows the gradient of an entire
training set downhill
• Stochastic gradient descent, follows the gradient of
randomly selected minbatches downhill
Monte Carlo discussions ….
Auto-differentiation discussions …
Regularization

 loss term + lambda regularization term

 Regularization as constraints for underdetermined system
 Eg. A^T A + alpha I, Solving linear homogeneous Ax = 0, ||Ax||^2 with ||x||^2=1
 Regularization for the stable uniqueness solution of ill-posed problems (no unique solution)
 Regularization as prior for MAP on top of the maximum likelihood estimation, again all
approximations to the full Bayesian inference

 L1/2 regularization
 L1 regularization has ‘feature selection’ property, that is, it produces a sparse vector,
setting many to zeros
 L2 is usually diffuse, producing small numbers.
 L2 superior is not explicitly selecting features
Architecture prior and inductive biases
discussions ...
Hyperparameters and validation
• The hyper-parameter lambda
• Split the training data into two disjoint subsets
– One is to learn the parameters
– The other subset, the validation set, is to guide the
selection of the hyperparameters
A linearly separable toy example

• The toy spiral data consists of three classes (blue, red, yellow) that are not
linearly separable.
– 300 pts, 3 classes
• Linear classifier fails to learn the toy spiral dataset.
• Neural Network classifier crushes the spiral dataset.
– One hidden layer of width 100
A toy example from cn231n

• The toy spiral data consists of three classes (blue, red, yellow) that are not
linearly separable.
– 3 classes, 100 pts for each class
• Softmax linear classifier fails to learn the toy spiral dataset.
– One layer, W, 2*3, b,
– analytical gradients, 190 iteration, loss from 1.09 to 0.78, 48% training set accuracy
• Neural Network classifier crushes the spiral dataset.
– One hidden layer of width 100, W1, 2*100, b1, W2, 100*3, only few extra lines of python codes!
– 9000 iteration, loss from 1.09 to 0.24, 98% training set accuracy
Generalization
• The ‘optimization’ reduces the training errors (or residual errors before)
• The ‘machine learning’ wants to reduce the generalization error or the
test error as well.

• The generalization error is the expected value of the error on a new input,
from the test set.
– Make the training error small.
– Make the gap between training and test error small.

• Underfitting is not able to have sufficiently low error on the training set
• Overfitting is not able to narrow the gap between the training and the
test error.
The training data was generated synthetically, by randomly sampling x values and choosing y deterministically
by evaluating a quadratic function. (Left) A linear function fit to the data suffers from
underfitting. (Center) A quadratic function fit to the data generalizes well to unseen points. It does not suffer from
a significant amount of overfitting or underfitting. (Right) A polynomial of degree 9 fit to
the data suffers from overfitting. Here we used the Moore-Penrose pseudoinverse to solve
the underdetermined normal equations. The solution passes through all of the training
points exactly, but we have not been lucky enough for it to extract the correct structure.
The capacity of a model

• The old Occam’s razor. Among competing ones, we should choose the
“simplest” one.
• The modern VC Vapnik-Chervonenkis dimension. The largest possible
value of m for which there exists a training set of m different x points that
the binary classifier can label arbitrarily.
• The no-free lunch theorem. No best learning algorithms, no best form of
regularization. Task specific.
Practice: learning rate

Left: A cartoon depicting the effects of different learning rates. With low learning rates the
improvements will be linear. With high learning rates they will start to look more exponential.
Higher learning rates will decay the loss faster, but they get stuck at worse values of loss
(green line). This is because there is too much "energy" in the optimization and the
parameters are bouncing around chaotically, unable to settle in a nice spot in the optimization
landscape. Right: An example of a typical loss function over time, while training a small
network on CIFAR-10 dataset. This loss function looks reasonable (it might indicate a slightly
too small learning rate based on its speed of decay, but it's hard to say), and also indicates
that the batch size might be a little too low (since the cost is a little too noisy).
Practice: avoid overfitting

The gap between the training and validation accuracy

indicates the amount of overfitting. Two possible cases
are shown in the diagram on the left. The blue validation
error curve shows very small validation accuracy
compared to the training accuracy, indicating strong
overfitting. When you see this in practice you probably
want to increase regularization or collect more data. The
other possible case is when the validation accuracy tracks
the training accuracy fairly well. This case indicates that
your model capacity is not high enough: make the model
larger by increasing the number of parameters.
 Why deeper?

• Deeper networks are able to use far fewer units per layer
and far fewer parameters, as well as frequently
generalizing to the test set
• But harder to optimize!
• Choosing a deep model encodes a very general belief that
the function we want to learn involves composition of
several simpler functions. Or the learning consists of
discovering a set of underlying factors of variation that can
in turn be described in terms of other, simpler underlying
factors of variation.
Curse of dimensionality

• The core idea in deep learning is that we assume that

the data was generated by the composition factors or
features, potentially at multiple levels in a hierarchy.
• This assumption allows an exponential gain in the
relationship between the number of examples and
the number of regions that can be distinguished.
• The exponential advantages conferred by the use of
deep, distributed representations counter the
exponential challenges posed by the curse of
dimensionality.
Convolutional Neural Networks, or CNN,
a visual network, and back to a 2D lattice from the abstraction of the 1D feature vector x in neural networks
From a regular network to CNN

• We have regarded an input image as a vector of features, x, by virtue of

feature detection and selection, like other machine learning applications to
learn f(x)
• We now regard it as is, a 2D image, a 2D grid, a topological discrete lattice,
I(I,j)

• Converting input images into feature vector loses the spatial neighborhood-
ness
• complexity increases to cubics
• Yet, the connectivities become local to reduce the complexity!
 What is a convolution?

• The fundamental operation, the convolution I * K(I,j) = \sum \

sum
• Flipping kernels makes the convolution commutative, which is
fundamental in theory, but not required in NN to compose
with other functions, so to include discrete correlations as well
into “convolution”
• Convolution is a linear operator, dot-product like correlation,
not a matrix multiplication, but can be implemented as a
sparse matrix multiplication, to be viewed as an affine
transform
 A regular 3-layer Neural Network.

A CNN arranges its neurons in three dimensions (width, height, depth). Every layer
of a CNN transforms the 3D input volume to a 3D output volume. In this example,
the red input layer holds the image, so its width and height would be the dimensions
of the image, and the depth would be 3 (Red, Green, Blue channels).
 Convolutional Neural Networks: 1998. Input 32*32. CPU

LeNet: a layered model composed of convolution and

subsampling operations followed by a holistic
representation and ultimately a classifier for
handwritten digits. [ LeNet ]
 Convolutional Neural Networks: 2012. Input 224*224*3. GPU.

AlexNet: a layered model composed of convolution, + data

subsampling, and further operations followed by a + gpu
holistic representation and all-in-all a landmark + non-saturating nonlinearity
classifier on + regularization
ILSVRC12. [ AlexNet ]
 LeNet vs AlexNet

• 224*224*3
• 32*32*1 • 8 layers
• 7 layers • 5 conv and 3 fully classification
• 2 conv and 4 classification • 5 convolutional layers, and 3,4,5 stacked on top o
• 60 thousand parameters each other
• Three complete conv layers
• Only two complete convolutional
layers • 60 million parameters, insufficient data
– Conv, nonlinearities, and pooling as
• Data augmentation:
one complete layer – Patches (224 from 256 input), translations, reflection
– PCA, simulate changes in intensity and colors
 The motivation of convolutions

• Sparse interaction, or Local connectivity.

– The receptive field of the neuron, or the filter size.
– The connections are local in space (width and height), but
always full in depth
– A set of learnable filters
• Parameters sharing, the weights are tied
• Equivariant representation, translation invariant
 Convolution and matrix multiplication

• Discrete convolution can be viewed as multiplication

by matrix
• The kernel is a doubly block circulant matrix
• It is very sparse!
 The ‘convolution’ operation
• The convolution is commutative because we have flipped the kernel
– Many implement a cross-correlation without flipping
• A convolution can be defined for 1, 2, 3, and N D
– The 2D convolution is different from a real 3D convolution, which integrates the
spatio-temporal information, the standard CNN convolution has only ‘spatial’
spreading
• In CNN, even for 3D RGB input images, the standard convolution is 2D in each
channel,
– each channel has a different filter or kernel, the convolution per channel is then
summed up in all channels to produce a scalar for non-linearity activation
– The filiter in each channel is not normalized, so no need to have different linear
combination coefficients.
– 1*1 convolution is a dot product in different channel, a linear combination of different
chanels
• The backward pass of a convolution is also a convolution with spatially flipped
filters.
VisGraph, HKUST
 The convolution layers

• Stacking several small convolution layers is different from

convolution cascating
– As each small convolution is followed by the nonlinearities ReLU
– The nonlinearities make the features more expressive!
– Have fewer parameters with small filters, but more memory.
– Cascating simply enlarges the spatial extent, the receptive field
• Whether each conv layer is also followed by a pooling?
– Lenet does not!
– AlexNet first did not.

VisGraph, HKUST
 The Pooling Layer

• Reduce the spatial size

• Reduce the amount of parameters
• Avoid over-fitting

• Backpropagation for a max: only routing the gradient to

the input that had the highest value in the forward pass

• It is unclear whether the pooling is essential.

VisGraph, HKUST
Pooling layer down-samples the volume spatially, independently in each depth slice of
the input volume.
Left: the input volume of size [224x224x64] is pooled with filter size 2, stride 2 into
output volume of size [112x112x64]. Notice that the volume depth is preserved.
Right: The most common down-sampling operation is max, giving rise to max
pooling, here shown with a stride of 2. That is, each max is taken over 4 numbers
(little 2x2 square).
 The spatial hyperparameters

• Depth
• Stride
• Zero-padding
 AlexNet 2012

• A strong prior has very low entropy, e.g. a Gaussian

with low variance
• An infinitely strong prior says that some parameters
are forbidden, and places zero probability on them
• The convolution ‘prior’ says the identical and zero
weights
• The pooling forces the invariance of small
translations
 The convolution and pooling act as an
infinitely strong prior!

• A strong prior has very low entropy, e.g. a Gaussian

with low variance
• An infinitely strong prior says that some parameters
are forbidden, and places zero probability on them
• The convolution ‘prior’ says the identical and zero
weights
• The pooling forces the invariance of small
translations
 The neuroscientific basis for CNN
• The primary visual cortex, V1, about which we know the most
• The brain region LGN, lateral geniculate nucleus, at the back of the head carries the
signal from the eye to V1, a convolutional layer captures three aspects of V1
– It has a 2-dimensional structure
– V1 contains many simple cells, linear units
– V1 has many complex cells, corresponding to features with shift invariance, similar to pooling
• When viewing an object, info flows from the retina, through LGN, to V1, then onward
to V2, then V4, then IT, inferotemporal cortex, corresponding to the last layer of CNN
features
• Not modeled at all. The mammalian vision system develops an attention mechanism
– The human eye is mostly very low resolution, except for a tiny patch fovea.
– The human brain makes several eye movements saccades to salient parts of the scene
– The human vision perceives 3D
• A simple cell responds to a specific spatial frequency of brightness in a specific
direction at a specific location --- Gabor-like functions
Receptive field

Left: An example input volume in red (e.g. a 32x32x3 CIFAR-10 image), and an example
volume of neurons in the first Convolutional layer. Each neuron in the convolutional layer is
connected only to a local region in the input volume spatially, but to the full depth (i.e. all
color channels). Note, there are multiple neurons (5 in this example) along the depth, all
looking at the same region in the input - see discussion of depth columns in text below.
Right: The neurons from the Neural Network chapter remain unchanged: They still
compute a dot product of their weights with the input followed by a non-linearity, but their
connectivity is now restricted to be local spatially.
Receptive field
Gabor functions
Gabor-like learned features
CNN architectures and algorithms
 CNN architectures
• The conventional linear structure, linear list of layers, feedforward
• Generally a DAG, directed acyclic graph
• ResNet simply adds back

• Different terminology: complex layer and simple layer

– A complex (complete) convolutional layer, including different stages such as
convolution per se, batch normalization, nonlinearity, and pooling
– Each stage is a layer, even there are no parameters
• The traditional CNNs are just a few complex convolutional layers to
extract features, then are followed by a softmax classification output
layer
• Convolutional networks output a high-dimensional, structured object,
rather than just predicting a class label for a classiciation task or a real
valuefor a regression task, it it an output tensor
– S_i,j,k is the probability that pixel j,k belongs to class I
 The popular CNN

• LeNet, 1998
• AlexNet, 2012
• VGGNet, 2014
• ResNet, 2015
 VGGNet

• 16 layers
• Only 3*3
convolutions
• 138 million
parameters
ResNet

• 152 layers
• ResNet50
 Computational complexity
• The memory bottleneck
• GPU, a few GB
Stochastic Gradient Descent

• Gradient descent, follows the gradient of an

entire training set downhill

• Stochastic gradient descent, follows the

gradient of randomly selected minbatches
downhill
The dropout regularization

• Randomly shutdown a subset of units in training

• It is a sparse representation
• It is a different net each time, but all nets share the parameters
– A net with n units can be seen as a collection of 2^n possible thinned nets,
all of which share weights.
– At test time, it is a single net with averaging
• Avoid over-fitting
 Ensemble methods

• Bagging (bootstrap aggregating), model averaging,

ensemble methods
– Model averaging outperforms, with increased computation
and memory
– Model averaging is discouraged when benchmarking for
publications
• Boosting, an ensemble method, constructs an
ensemble with higher capacity than the individual
models
 Batch normalization

• After convolution, before nonlinearity

• ‘Batch’ as it is done for a subset of data
• Instead of 0-1 (zero mean unit variance) input, the
distribution will be learnt, even undone if necessary

• The data normalization or PCA/whitening is common

in general NN, but in CNN, the ‘normalization layer’
has been shown to be minimal in some nets as well.
Open questions

• Networks are non-convex

– Need regularization
• Smaller networks are hard to train with local
methods
– Local minima are bad, in loss, not stable, large variance
• Bigger ones are easier
– More local minima, but better, more stable, small variance
• As big as the computational power, and data!
 Local minima, saddle points, and
plateaus?
• We don’t care about finding the exact mimimum, we only
want to obtain good generalization error by reducing the
function value.

• In low-dimensional, local minima are common.

• In higher dimensional, local minima are rare, and saddle
points are more common.
– The Hessian at a local mimimum has all positive eigenvalues. The
Hessian at a saddle point has a micture of pos and neg
eigenvalues.
– It is exponentially unlikely that all n will have the same sign in
high n-dim
CNN applications
 CNN applications

• Transfer learning
• Fine-tuning the CNN
– Keep some early layers
• Early layers contain more generic features, edges, color blobs
• Common to many visual tasks
– Fine-tune the later layers
• More specific to the details of the class
• CNN as feature extractor
– Remove the last fully connected layer
– A kind of descriptor or CNN codes for the image
– AlexNet gives a 4096 Dim descriptor
 CNN classification/recognition nets

• CNN layers and fully-connected classification layers

• From ResNet to DenseNet
– Densely connected
– Feature concatenation

VisGraph, HKUST
 Fully convolutional nets: semantic
segmentation
• Classification/recognition nets produce ‘non-spatial’
outputs
– the last fully connected layer has the fixed dimension of
classes, throws away spatial coordinates

• Fully convolutional nets output maps as well

Semantic segmentation
Using sliding windows for semantic
segmentation
Fully convolutional
Fully convolutional
 Detection and segmentation nets: The
Mask Region-based CNN (R-CNN):
• Class-independent region (bounding box) proposals
– From selective search to region proposal net with objectness
• Use CNN to class each region
• Regression on the bounding box or contour segmentation

• Mask R-CNN: end-to-end

– Use CNN to make proposals on object/non-object in parallel

• The old good idea of face detection by Viola

– Proposal generation
– Cascading (ada boosting)

VisGraph, HKUST
Using sliding windows for object
detection as classification
Mask R-CNN
Excellent results
End.
Some old notes in 2017 and 2018

VisGraph, HKUST
 Fundamentally from continuous to discrete
views … from geometry to recognition
• ‘Simple’ neighborhood from topology
• discrete high order

• Modeling higher-order discrete, but yet solving it with

the first-order differentiable optimization
• Modeling and implementation become easier

• The multiscale local jet, a hierarchical and local

characterization of the image in a full scale-space
neighborhood
 Local jets? Why?

• The multiscale local jet, a hierarchical and local

characterization of the image in a full scale-space
neighborhood

• A feature in CNN is one element of the descriptor

VisGraph, HKUST
 Classification vs regression

• Regression predicts a value from a continuous set, a

real number/continuous,
– Given a set of data, find the relationship (often, the
continuous mathematical functions) that represent the set
of data
– To predict the output value using training data
• Whereas classification predicts the ‘belonging’ to the
class, a discrete/categorial variable
– Given a set of data and classes, identify the class that the
data belongs to
– To group the output into a class
 Autoencoder and decoder

• Compression, and reconstruction

• Convolutional autoencoder, trained to reconstruct its
input
– Wide and thin (RGB) to narrow and thick
• ….
 Convolution and deconvolution

• The convolution is not inversible, so there is no strict

definition, or the closed-form of the so-called
‘deconvolution’
• In iterative procedure, a kind of the convolution
transpose is applied, so to call it ‘ deconvolution’
• The ‘deconvolution’ filters are reconstruction bases
 CNN as a natural features and
descriptors
• Each point is an interest point or feature point with
high-dimensional descriptors
• Some of them are naturally ‘unimportant’, and
weighted down by the nets
• The spatial information is kept through the nets
• The hierarchical description is natural from local to
global for each point, each pixel, and each feature
point
Traditional stereo vs deep stereo
regression
 CNN regression nets: deep regression
stereo
• Regularize the cost volume.

VisGraph, HKUST
 Traditional stereo
• Input image H * W * C

• (The matching) cost volume in disparities D, or in depths

• D*H*W
• The value d_i for each D is the matching cost, or the
correlation score, or the accumulated in the scale space, for
the disparity i, or the depth i.
• Disparities are a function of H and W, d = c(x,y;x+d,y+d)

• Argmin D
• H*W
 End-to-end deep stereo regression
• Input image H * W * C
• 18 CNN
• H*W*F

• (F, features, are descriptor vectors for each pixel, we may just correlate or dot-product two descriptor
vectors f1 and f2 to produce a score in D*H*W. But F could be further redefined in successive convolution
layers.)

• Cost volume, for each disparity level

• D * H * W * 2F
• 4D volume, viewed as a descriptor vector 2F for each voxel D*H*W
• 3D convolution on H, W, and D
• 19-37 3D CNN

• The last one (deconvolution) turns F into a scalar as a score

• D*H *W

• Soft argmin D
• H*W
Bayesian decision

 posterior = likelihood * prior / evidence

 Decide if > ; otherwise decide
Statistical measurements

 Precision
 Recall
 coverage
 Reinforcement learning (RL)

• Dynamic programming uses a mathematical model of

the environment, which is typically formulated as a
Markov Decision Process
• The main difference between the classical dynamic
programming and RL is that RL does not assume an
exact mathematical model of the Markov Decision
Process, and target large MDP where exact methods
become infeasible
• Therefore, RL is neuro dynamic programming. In
operational research and control, it’s called approximate
dynamic programming
VisGraph, HKUST
Non-linear iterative optimisation

• J d = r from vector F(x+d)=F(x)+J d

• minimize the square of y-F(x+d)=y-F(x)-J d = r – J d
• normal equation is J^T J d = J^T r (Gauss-Newton)
• (H+lambda I) d = J^T r (LM)

Note: F is a vector of functions, i.e. min f=(y-F)^T(y-F)

116
 General non-linear optimisation

• 1-order , d gradient descent d= g and H =I

• 2-order,
• Newton step: H d = -g
• Gauss-Newton for LS: f=(y-F)^T(y-F), H=J^TJ, g=-J^T r
• ‘restricted step’, trust-region, LM: (H+lambda W) d = -g

R. Fletcher: practical method of optimisation

f(x+d) = f(x)+g^T d + ½ d^T H d
Note: f is a scalar valued function here.
117
statistics

• ‘small errors’ --- classical estimation theory

• analytical based on first order appoxi.
• Monte Carlo

• ‘big errors’ --- robust stat.

Talk about it later
• RANSAC
• LMS
• M-estimators

118
An abstract view
 The input
 The classification with linear models,
 Can be a SVM
 The output layer of the linear softmax classifier
 find a map or a transform to make them linear, but in higher
dimensions
 provides a set of features describing
 Or provides a new representation for
 The nonlinear transformation
 Can be hand designed kernels in svm
 It is the hidden layer of a feedforward network
 Deep learning
 To learn
 is compositional, multiple layers
 CNN is convolutional for each layer
 To do
• Add drawing for ‘receptive fields’
• Dot product for vectors, convolution, more specific for 2D? Time
invariant, or translation invariant, equivariant
• Convolution, then nonlinear acitivation, also called ‘detection
stage’, detector
• Pooling is for efficiency, down-sampling, and handling inputs of
variable sizes,

VisGraph, HKUST
Cost or loss functions - bis

• Classification and regression are different! for different application communities

• We used more traditional regression, but the modern one is more on classification, so
resulting different loss considerations
• Regression is harder to optimize, while classification is easier, and more stable

• Classification is easier than regression, so always discretize and quantize the output, and
convert them into classification tasks!

• One big improvement in NN modern development is that the cross-entropy dominates the
mean squred error L2, as the mean squared error was popular and good for regression, but
not that good for NN. because of its fundamental more appropriate distribution assumption,
not normal distributions

• L2 for regression is harder to optimize than the more stable softmax for classification
• L2 is also less robust
Automatic differentiation (algorithmic diff),
and backprop, its role in the development
• Differentiation: symbolic or numerical (finite differences)
• Automatic differentiation is to compute derivatives algorithmically,
backprop is only one approach to it.
• Its history is related to that of NN and deep learning

• Worked for traditional small systems, a f(x)

• Larger and more explicit composional nature of f1(f2(f3(… fn(x))))
goes back to the very nature of the derivatives
– Forward mode and reverse mode (it is based on f(a+b epsilon) =f(a)+f’(a)b
epsilon, f(g(a+b epsilon) = f(g(a))+f’(g(a))g’(a) b epsilon)
– The reverse mode is backprop for NN
• In the end, it benefits as well the classical large optimization such as
bundle adjustment
Viewing the composition of an
arbitrary function as a natural layering
• Take f(x,y)=x+s(y) / s(x) + (x+y)^2, at a given point x=3,y=-4
• Forward pass
• f1=s(y), f2 = x+f1, f3=s(x), f4=x+y, f5=f4^2, f6=f3+f5, f7=1/f6, f8=f2*f7
• So f(*)=f8(f7(f6(f5(f4(f3(f2(f1(*)))))))), each of fn is a known elementary function or
operation

• Backprop to get (df/dx, df/dy), or abreviated as (dx,dy), at (3,-4)

• f8=f, abreviate df/df7 or df8/df7 as df7, df/df7=f2,…, and df/dx as dx, …
• df7=f2, (df2=f7), df6= (-1/f6^2) * df7, df5=df6, (df3=df6), df4=(2*f4)df5, dx=df4, dy=df4,
dx += (1-s(x)*s(x)*df3 (backprop in s(x)=f3), dx += df2 (backprop in f2), dy += df2
(backprop in f2), dy += (1-s(y))*s(y)*df1 (backprop in s(y)=f1)

• In NN, there are just more variables in each layer, but the elementary functions are
much simpler: add, multiply, and max.
• Even the primitive function in each layer takes also the simplest one! Then just a lot of
them!
Computational graph

• NN is described with a relatively informal graph

language
• More precise computational graph to describe the
backprop algorithms
Structured probabilistic models,
Graphical models, and factor graph
• A structured probabilistic model is a way of
describing a probability distribution.
• Structured probabilistic models are referred to as
graphical models
• Factor graphs are another way of drawing undirected
models that resolve an ambiguity in the graphical
representation of standard undirected model syntax.