S.No.
|
Topics
|
Lectures
|
Instructor
|
References/Notes
|
0
|
Introduction to Machine Learning
|
01-01
|
SDR
|
|
|
Flavours of Machine Learning: Unsupervised, Supervised,
Reinforcement, Hybrid models. Decision Boundaries: crisp, and
non-crisp, optimisation problems. Examples of unsupervised learning.
|
03 Jan (Tue) {lecture#01}
|
SDR
|
[Online-only class: MS-Teams: 07:00am-08:00am]
MS-Teams folder:
slides_k_means_em1_03jan23.pdf,
video_k_means_em1_03jan23.mp4
|
1
|
Unsupervised Learning:
K-Means, Gaussian Mixture Models, EM
|
01-06
|
SDR
|
[Bishop Chap.9],
[Do: Gaussians],
[Do: More on Gaussians],
[Ng: K-Means],
[Ng: GMM],
[Ng: EM],
[Smyth:
EM]
|
|
The K-Means algorithm: Introduction. Algorithms: history,
flavours. A mathematical formulation of the K-Means algorithm.
The Objective function to minimise.
The K-Means algorithm:
The Objective function to minimise.
|
03 Jan (Tue) {lecture#01}
|
SDR
|
MS-Teams folder:
slides_k_means_em1_03jan23.pdf,
video_k_means_em1_03jan23.mp4
|
|
The basic K-Means algorithm, computation complexity issues: each
step, overall. Limitations of K-Means.
|
04 Jan (Wed) {lecture#02}
|
SDR
|
MS-Teams folder:
video_k_means_em2_04jan23.mp4,
slides_k_means_em2_04jan23.pdf
|
|
Limitations of K-Means.
K-Means: Alternate formulation with a distance threshold.
An introduction to Gaussian Mixture Models.
|
06 Jan (Fri) {lecture#03}
|
SDR
|
MS-Teams folder:
video_k_means_em3_06jan23.mp4,
slides_k_means_em3_06jan23.pdf,
lecture_notes_k_means_em3_06jan23.pdf
|
|
The Bayes rule, and Responsibilities.
Maximum Likelihood Estimation. Parameter estimation for a mixture
of Gaussians, starting with a simple 1-D single Gaussian case.
ML-Estimation: the simple case of one 1-D Gaussian, to
the general case of K D-dimensional Gaussians.
The Mahalanobis Distance.
|
10 Jan (Tue) {lecture#04}
|
SDR
|
MS-Teams folder:
video_k_means_em4_10jan23.mp4,
slides_k_means_em4_10jan23.pdf
|
|
The general case of K D-dimensional Gaussians
(contd).
Getting stuck, using Lagrange Multipliers.
The EM Algorithm for Gaussian Mixtures.
Application: Assignment 1:
The Stauffer and Grimson Adaptive Background Subtraction
Algorithm.
An introduction to the basic set of interesting heuristics!
|
11 Jan (Wed) {lecture#05}
|
SDR
|
MS-Teams folder:
video_k_means_em5_11jan23.mp4,
slides_k_means_em5_11jan23.pdf
|
|
The Stauffer and Grimson algorithm (contd)
|
13 Jan (Fri) {lecture#06}
|
SDR
|
MS-Teams folder:
|
2
|
Unsupervised Learning:
EigenAnalysis:
PCA, LDA and Subspaces
|
06-10
|
SDR
|
[Ng: PCA],
[Ng: ICA],
[Burges: Dimension Reduction],
[Bishop Chap.12]
|
|
Introduction to Eigenvalues and Eigenvectors.
Properties of Eigenvalues and Eigenvectors.
|
17 Jan (Tue) {lecture#07}
|
SDR
|
MS-Teams folder:
video_eigen1_17jan23.mp4, slides_eigen1_17jan23.pdf
|
|
Properties of Eigenvalues and Eigenvectors (contd).
Gram-Schmidt Orthogonalisation, other properties.
The KL Transform (contd).
The SVD and its properties.
|
18 Jan (Wed) {lecture#08}
|
SDR
|
MS-Teams folder:
video_eigen2_18jan23.mp4, slides_eigen2_18jan23.pdf
|
|
The SVD and its properties (contd).
Application: Assignment 2: Eigenfaces and Fisherfaces
|
20 Jan (Fri) {lecture#09}
|
SDR
|
MS-Teams folder:
video_eigen3_linear1_20jan23.mp4,
slides_eigen3_20jan23.pdf
|
3
|
Linear Models for Regression, Classification
|
10-14
|
SDR
|
[Bishop Chap.3],
[Bishop Chap.4],
[Ng: Supervised, Discriminant Analysis],
[Ng: Generative]
|
|
General introduction to Regression and Classification.
|
20 Jan (Fri) {lecture#09}
|
SDR
|
MS-Teams folder:
video_eigen3_linear1_20jan23.mp4,
slides_linear1_20jan23.pdf
|
|
Linearity and restricted non-linearity.
Maximum Likelihood and Least Squares.
The Moore-Penrose Pseudo-inverse.
Regularised Least Squares.
Three approaches to classification: restricted non-linear models
(linear combination of possible non-linear feature transformations).
Introduction to linear models: equation of a line in terms of the
physical significance of the space, and the weights
w.
|
24 Jan (Tue) {lecture#10}
|
SDR
|
MS-Teams folder:
video_linear2_24jan23.mp4,
slides_linear2_24jan23.pdf,
lecture_notes_linear2_24jan23.pdf
|
|
Introduction to linear models: equation of a line in terms of the
physical significance of the space, and the weights
w (contd).
Linear Discriminant Functions: 2 classes, and K classes. Fisher's
Linear Discriminant (basic build-up).
|
25 Jan (Wed) {lecture#11}
|
SDR
|
MS-Teams folder:
video_linear3_25jan23.mp4,
slides_linear3_25jan23.pdf
|
|
Fisher's Linear Discriminant.
Application: Assignment 2: Eigenfaces and Fisherfaces
|
26 Jan (Thu) {lecture#12}
|
SDR
|
[Online-only class: MS-Teams: 07:00am-08:00am]
(Make-up lecture for the missed 27 Jan (Fri) lecture)
MS-Teams folder:
video_linear4_svm1_26jan23.mp4,
slides_linear4_26jan23.pdf,
|
|
---
|
27 Jan (Fri) {lecture#xx}
|
SDR
|
(No class)
|
4
|
SVMs and Kernels
|
12-18
|
SDR
|
[Bishop Chap.7],
[Alex: SVMs],
[Ng: SVMs],
[Burges: SVMs],
[Bishop Chap.6]
|
|
SVMs: the concept of the margin.
|
26 Jan (Thu) {lecture#12}
|
SDR
|
[Online-only class: MS-Teams: 07:00am-08:00am]
(Make-up lecture for the missed 27 Jan (Fri) lecture)
MS-Teams folder:
video_linear4_svm1_26jan23.mp4,
slides_svm1_26jan23.pdf
|
|
SVMs: the optimisation problem, getting
the physical significance of the y = +1 and y = -1 lines.
The two `golden' regions for the 2-class perfectly separable case.
The generalised canonical representation in terms of one inequation.
The basic SVM optimisation: the primal and the dual problems.
An illustration of the kernel trick.
|
29 Jan (Sun) {lecture#13}
|
SDR
|
[Online-only class: MS-Teams: 07:30pm-08:30pm]
(Make-up lecture for the missed 03 Feb (Fri) lecture)
MS-Teams folder:
video_svm2_29jan23.mp4,
slides_svm2_29jan23.pdf
|
|
Lagrange Multipliers and the KKT Conditions.
An illustration of the kernel trick.
|
31 Jan (Tue) {lecture#14}
|
SDR
|
MS-Teams folder:
video_svm3_31jan23.mp4,
slides_svm3_31jan23.pdf
|
|
Lagrange Multipliers and the KKT Conditions (contd).
The Soft-Margin SVM.
|
01 Feb (Wed) {lecture#15}
|
SDR
|
MS-Teams folder:
video_svm4_01feb23.mp4,
slides_svm4_01feb23.pdf
|
|
---
|
03 Feb (Fri) {lecture#xx}
|
SDR
|
(No class)
|
---
|
Minor-1
|
Minor-1: 07 Feb (Tue), 08:00am-09:00am, LH-408
|
---
|
|
|
The Soft-Margin SVM (contd).
Abstracting the basic concepts of the
hard-margin SVM, to use in a similar formulation. The function to
optimise, the inequality constraints, the KKT conditions from
Lagrange's theory. The Primal and Dual formulations.
Lagrange Multipliers and the KKT Conditions.
Introduction to Kernels.
|
10 Feb (Fri) {lecture#16}
|
SDR
|
MS-Teams folder:
video_svm5_kernel1_10feb23.mp4,
slides_svm5_kernel1_10feb23.pdf,
lecture_notes_kernel1_10feb23.pdf
|
|
Introduction to Kernels. Kernels in Regression.
|
14 Feb (Tue) {lecture#17}
|
SDR
|
MS-Teams folder:
video_kernel2_14feb23.mp4,
lecture_notes_kernel2_14feb23.pdf
|
|
Kernels: construction, properties, examples.
|
15 Feb (Wed) {lecture#18}
|
SDR
|
MS-Teams folder:
video_kernel3_nn1_15feb23.mp4,
lecture_notes_kernel3_15feb23.pdf
|
5
|
Neural Networks
|
18-41
|
SDR
|
[Bishop Chap.5], [DL Chap.6], [DL Chap.9]
|
|
Introduction to Neural Networks. Perceptron: a linear classifier.
A non-neural network interpretation.
|
15 Feb (Wed) {lecture#18}
|
SDR
|
MS-Teams folder:
video_kernel3_nn1_15feb23.mp4,
slides_nn1_15feb23.pdf
|
|
The Perceptron: a neural interpretation. The Perceptron
optimisation. Weight update. Conventions of the Multi-layer
Perceptron (MLP). The X-OR problem with the Perceptron.
|
17 Feb (Fri) {lecture#19}
|
SDR
|
MS-Teams folder:
video_nn2_17feb23.mp4,
lecture_notes_nn2_17feb23.pdf,
slides_nn2_17feb23.pdf
|
|
The XOR problem with the Perceptron with a kernel function.
Examples of neural network activation functions.
Factorisation: probability, differential calculus.
|
21 Feb (Tue) {lecture#20}
|
SDR
|
MS-Teams folder:
video_nn3_21feb23.mp4,
slides_nn3_21feb23.pdf,
lecture_notes_nn3_21feb23.pdf
|
|
Pre-backpropagation mathematical fundamentals (contd).
Factorisation, and the chain rule in differential calculus.
The Perceptron and the Multi-Layer Perceptron (MLP).
|
22 Feb (Wed) {lecture#21}
|
SDR
|
MS-Teams folder:
video_nn4_22feb23.mp4,
lecture_notes_nn4_22feb23.pdf
|
|
The BackPropagation Algorithm: a computational mechanism for the
chain rule in differential calculus.
|
28 Feb (Tue) {lecture#22}
|
SDR
|
MS-Teams folder:
video_nn5_28feb23.mp4,
lecture_notes_nn5_28feb23.pdf
|
|
Interesting interpetations of MLPs.
Three ways to have kernel functions/feature trasnformations.
The interpretation of a hidden layer of an MLP as a kernel
function or a (possibly non-linear) feature transformation.
Trying the XOR problem as a regression problem. It will not work,
as expected. Now, trying this with a hand-crafted example of an
MLP with hand-crafted weights: a single hidden layer with 2
neurons, and the hidden layer having the ReLU activation function.
Emboldened by this, we try implementing other "basic" Boolean
functions with handcrafted neural networks. Attempts with a
traditional sigmoid activation function. Getting some hands-on
feel with weight values and sigmoid-specific constraints for
ranges of inputs. The NOT and AND functions.
|
01 Mar (Wed) {lecture#23}
|
SDR
|
MS-Teams folder:
video_nn6_01mar23.mp4,
lecture_notes_nn6_01mar23.pdf
|
|
Other "basic" Boolean functions with hand-crafted neural networks (contd).
From MLPs, a few steps towards deep neural networks. Inputs as images.
The interpretation of a weight vector as an image. Empirical
observations regarding the initial layers of neural networks, and
the overall number of connections. The initial layers compute
directional derivatives. The overall connections are such that
only a small number of local connections are effectively there,
in the neighbourhood of a neuron.
|
03 Mar (Fri) {lecture#24}
|
SDR
|
MS-Teams folder:
video_nn7_03mar23.mp4,
lecture_notes_nn7_03mar23.pdf
|
|
Some insight into 2-D inputs, and intepreting first layer weights
as images, and its importance for CNNs (contd): an empirical
observation, and evidence for the idea of local receptive fields,
low-level differentiation/edge features, with some biological
motivation as well, from the visual cortex.
The empirical result about the interpretation of weights in the
first layer of a neural network with many layers.
Yet another example of an XOR implementation (semi-hand-crafted),
with a different architecture, different activation function, and
different inputs. The minterm connotation of the hidden layer.
iSome insight into the expressive power of feed-forward neural
networks: the insight from shallow networks with a large width.
An example with asymmetric values for inputs and outputs for a
digital circuit, and estimating neural network parameters, for D
input neurons and 2^D neurons in one hidden layer, and one output
neuron.
Convolution and Correlation: A domain-independent introduction.
Linear Shift/Time Invariant Systems
|
14 Mar (Tue) {lecture#25}
|
SDR
|
MS-Teams folder:
video_nn8_14mar23.mp4,
lecture_notes_nn8_14mar23.pdf
|
|
Linear Shift/Time Invariant Systems: characterisation of a system
in terms of a `standard' input (the single impulse, the Kronecker
Delta for instance, in the discrete domain, and the Dirac Delta,
in the continuous domain). The impulse response gives a
hardware-invariant characterisation of a system. Examples from
Electrical and Civil engineering.
Why is Convolution so fundamental in Linear Shift/Time Invariant systems?
A graphical proof.
Introduction to invariance in neural networks: translational
invariance (`Where's Wally?' or `Where's Waldo?').
|
15 Mar (Wed) {lecture#26}
|
SDR
|
MS-Teams folder:
video_nn9_15mar23.mp4,
lecture_notes_nn9_15mar23.pdf
|
|
Some characteristics of deep neural networks:
local receptive fields, strided convolutions, parameter sharing,
pooling.
Two basic deep neural network architectures: the contractive
structure, and the bow-tie structure.
Introduction to LeNet-5 (1989), the first successful basic deep neural
network architecture. A basic contractive architecture.
Illustration of the basic operations in the first two layers.
Introduction to an Auto-encoder: a basic bow-tie structure.
|
17 Mar (Fri) {lecture#27}
|
SDR
|
MS-Teams folder:
video_nn10_17mar23.mp4,
lecture_notes_nn10_17mar23.pdf
|
|
Autoencoder basics: (contd.)
Summarising a contractive architecture: convolution, pooling and
finally, a fully connected layer. A timeline of deep networks and
trends.
An example of the first successful contractive architecture,
LeNet-5 (1989).
|
19 Mar (Sun) {lecture#28}
|
SDR
|
[Online-only class: MS-Teams: 08:00am-09:00am]
(Make-up lecture for the missed 31 Mar (Fri) lecture)
MS-Teams folder:
video_nn11_19mar23.mp4,
lecture_notes_nn11_19mar23.pdf
|
|
LeNet-5 detailed description (contd).
Layers C1, S2, the weird C3 with the strange asymmetric mapping.
|
21 Mar (Tue) {lecture#29}
|
SDR
|
MS-Teams folder:
video_nn12_21mar23.mp4,
lecture_notes_nn12_21mar23.pdf
|
|
LeNet-5 detailed description (contd). Completing the LeNet-5
layers: including the fully connected layers at the end, and the
SoftMax activation function in the output.
An introduction to some deep learning concepts prior to exploring
the second successful representative architectures, AlexNet.
Pooling. Local Response Normalisation.
|
22 Mar (Wed) {lecture#30}
|
SDR
|
MS-Teams folder:
video_nn13_22mar23.mp4,
lecture_notes_nn13_22mar23.pdf
|
---
|
Minor-2
|
Minor-2: 24 Mar (Fri), 07:30am-08:30am, LH-408
|
---
|
|
|
An introduction to some deep learning concepts (contd):
some concepts in deep networks.
Batch Normalisation,
Residual/Skip/Highway Connections
|
28 Mar (Tue) {lecture#31}
|
SDR
|
MS-Teams folder:
video_nn14_28mar23.mp4,
lecture_notes_nn14_28mar23.pdf
|
|
An introduction to some deep learning concepts (contd): some
concepts in deep networks.
Residual/Skip/Highway Connections, Dropout.
A detailed description of the second class of successful deep
architectures, AlexNet.
|
30 Mar (Thu) {lecture#32}
|
SDR
|
[Online-only class: MS-Teams: 07:00am-08:00am]
(Make-up lecture for the missed 01 Apr (Sat) lecture)
MS-Teams folder:
video_nn15_30mar23.mp4,
lecture_notes_nn15_30mar23.pdf
|
|
|
31 Mar (Fri) {lecture#xx}
|
SDR
|
(No class)
|
|
|
01 Apr (Sat) {lecture#xx}
|
SDR
|
(No class)
|
|
AlexNet (contd).
|
05 Apr (Wed) {lecture#33}
|
SDR
|
MS-Teams folder:
video_nn16_05apr23.mp4,
lecture_notes_nn16_05apr23.pdf
|
|
AlexNet (contd).
The VGG family: VGG-16 and VGG-19.
|
11 Apr (Tue) {lecture#34}
|
SDR
|
MS-Teams folder:
video_nn17_11apr23.mp4,
lecture_notes_nn17_11apr23.pdf
|
|
The VGG family: VGG-16 and VGG-19 (contd): the baisc idea of
small 3x3 convolutions to replace larger convolutions
while keeping a small parameter count.
The ResNet family: the main points have been covered before, in
the discussion on residual conenctions and their use.
Introduction to the Inception architecture and its facets.
Carrying over the small convolution kernel idea from
VGG-16/VGG-19. The significance of concatenating scale
information from the input to get a possibly larger output size.
Asymmetric convolutions.
|
12 Apr (Wed) {lecture#35}
|
SDR
|
MS-Teams folder:
video_nn18_12apr23.mp4,
lecture_notes_nn18_12apr23.pdf
|
|
The Inception architecture (contd): the use of asymmetric
convolutions, and feature concatenation (to get richer
information in terms of scale). The physical significance of
1x1 convolutions.
Recurrent Neural Networks: introduction in terms of hardware
saving, advantage in terms of modelling sequential processes, and
inputs which are not of a fixed size, and disadvantage in terms
of the short-term memory problem: backpropagation in time with
the vanishing gradient problem.
|
14 Apr (Fri) {lecture#36}
|
SDR
|
[Online-only class: MS-Teams: 08:00am-09:00am]
(Make-up lecture for the missed 28 Apr (Fri) lecture)
MS-Teams folder:
video_nn19_14apr23.mp4,
lecture_notes_nn19_14apr23.pdf
|
|
Recurrent Neural Networks (contd.): more on the `backpropagation
through time' concept, leading to `Short-Term Memory'.
RNNs: dealing with text inputs and outputs: 1-hot-encoding.
RNNs: what situations can an RNN model?
An introduction to the two ways out: LSTMs and GRUs. LSTM cell basics.
|
18 Apr (Tue) {lecture#37}
|
SDR
|
MS-Teams folder:
video_nn20_18apr23.mp4,
lecture_notes_nn20_18apr23.pdf
|
|
LSTM cell basics (contd).
|
19 Apr (Wed) {lecture#38}
|
SDR
|
MS-Teams folder:
video_nn21_19apr23.mp4,
lecture_notes_nn21_19apr23.pdf
|
|
|
21 Apr (Fri) {lecture#xx}
|
SDR
|
(No class)
|
|
LSTMs and GRUs (contd.)
Region CNNs (R-CNNs): basics
|
23 Apr (Sun) {lecture#39}
|
SDR
|
[Online-only class: MS-Teams: 08:00am-09:00am]
(Make-up lecture for the missed 29 Apr (Sat) lecture)
MS-Teams folder:
video_nn22_23apr23.mp4,
lecture_notes_nn22_23apr23.pdf
|
|
R-CNNs (contd). A bit on the Viola-Jones Face detector (a bit of
history on filters applied to regions of images).
|
25 Apr (Tue) {lecture#40}
|
SDR
|
MS-Teams folder:
video_nn23_25apr23.mp4,
lecture_notes_nn23_25apr23.pdf
|
|
RNNs: an example with categorical data with different feature
vector lengths.
|
26 Apr (Wed) {lecture#41}
|
SDR
|
MS-Teams folder:
video_nn24_26apr23.mp4,
lecture_notes_nn24_26apr23.pdf
|
|
|
28 Apr (Fri) {lecture#xx}
|
SDR
|
(No class)
|
|
|
29 Apr (Sat) {lecture#xx}
|
SDR
|
(No class)
|
---
|
Major
|
09 May (Tue), 08:00am-09:00am, LH-408
|
---
|
---
|
xx
|
Mathematical Basics for Machine Learning
|
xx-xx
|
xx
|
[Burges: Math for ML],
[Do,
Kolter: Linear Algebra Notes],
|