APL 745

Optimization - I
Dr. Rajdip Nayek
Block V, Room 418D
Department of Applied Mechanics
Indian Institute of Technology Delhi

E-mail: [email protected]
Course website: https://www.csccm.in/courses/deep-learning-for-mechanics
Key components of any ML algorithm

1. The data that we learn from

2. A model with parameters to predict the output

3. An loss function that quantifies how bad the model is doing

4. An optimization algorithm to adjust the modelʼs parameters to

minimize the loss function

2
Overview
• Training examples: 𝒙(1) , 𝑡 1 , 𝒙(2) , 𝑡 2 , ⋯ , 𝒙(𝑁) , 𝑡 𝑁

• Model: 𝑦 = ℎ𝜽 𝒙 , we accumulate all weights and biases into a vector 𝜽

• Loss function → Average over individual training losses

𝑁 𝑁
1 𝑖
1
ℒ = ෍ℓ = ෍ℓ 𝑦 𝑖 ,𝑡 𝑖
𝑁 𝑁
𝑖=1 𝑖=1
• Gradient of loss function w.r.t parameter 𝜽
𝑁
1 𝑖
∇𝜽 ℒ = ∇𝜽 ෍ ℓ
𝑁
𝑖=1
𝑁
1 𝑖
= ෍ ∇𝜽 ℓ
𝑁
𝑖=1

• In the last lecture, we talked about how to compute gradients ∇𝜽 ℓ 𝑖 of the

individual loss functions w.r.t. parameter vector 𝜽 using backprop
• In this lecture, we will talk about optimization of neural nets
3
Basics: Visualizing a loss surface
• When we train a neural network, we’re trying to minimize some loss function
ℒ, which is a function of the network’s parameters, which we’ll denote with
the vector 𝜽
• 𝜽 contains all the neural network weights and biases
• Suppose 𝜽 consists of two weights, 𝑤1 and 𝑤2 . Let’s visualize a loss surface ℒ

𝑤2
𝑤1 4
Basics: Visualizing a loss surface
• Minima are points that minimize the loss function within a small neighbourhood
• One of them is a local minimum
• Another one is a global minimum, a point which achieves the minimum loss
over all values of 𝜽

𝑤2
𝑤1
5
Basics: Contour plots

• Visualizing a surface in 3D can sometimes become a bit cumbersome

• Can we do a 2D visualization of the loss surface?

• Yes, let’s take a look at something known as contours

6
Basics: Contour plots

• Each of the contours is associated with a set of parameters where the loss
function takes a particular value
• A small distance between the contours indicates a steep slope along
that direction
• A large distance between the contours indicates a gentle slope along that
direction

7
Basics: Contour plots
𝑤2

𝑤1

• Can’t determine the magnitude of the gradient from the contour plot
• Can determine the gradient direction: the gradient is always orthogonal
(perpendicular) to the contours
• Recall: Gradient descent would happen in opposite direction of gradient
8
Overview
• Training examples: 𝒙(1) , 𝑡 1 , 𝒙(2) , 𝑡 2 , ⋯ , 𝒙(𝑁) , 𝑡 𝑁

• Model: 𝑦 = ℎ𝜽 𝒙 , we accumulate all weights and biases into a vector 𝜽

• Loss function → Average over individual training losses

𝑁 𝑁
1 𝑖
1
ℒ = ෍ℓ = ෍ℓ 𝑦 𝑖 ,𝑡 𝑖
𝑁 𝑁
𝑖=1 𝑖=1
• Gradient of loss function w.r.t parameter 𝜽
𝑁
1 𝑖
∇𝜽 ℒ = ∇𝜽 ෍ ℓ
𝑁
𝑖=1
𝑁
1 𝑖
= ෍ ∇𝜽 ℓ
𝑁
𝑖=1

• In the last lecture, we talked about how to compute gradients ∇𝜽 ℓ 𝑖 of the

individual loss functions w.r.t. parameter vector 𝜽 using backprop
• In this lecture, we will talk about optimization of neural nets
9
Batch Gradient Descent
𝑁
1
∇𝜽 ℒ = ෍ ∇𝜽 ℓ 𝑖 𝜽𝑛𝑒𝑤 ← 𝜽𝑜𝑙𝑑 − 𝜂 ∇𝜽 ℒ ቚ
𝑁 𝜽= 𝜽𝑜𝑙𝑑
𝑖=1

• Compute the total gradient ∇𝜽 ℒ by averaging over all individual gradients for
every training example, and then update the parameters
• The algorithm goes over the entire data once before updating the parameters

• This is known as batch gradient descent (BGD), since we treat the entire training
set as a batch

• Pros: There is no approximation. Each update

step guarantees that the loss will decrease

• Cons: However, BGD can be very time-

consuming for a large dataset

10
Batch Gradient Descent
𝑁
1
∇𝜽 ℒ = ෍ ∇𝜽 ℓ 𝑖 𝜽𝑛𝑒𝑤 ← 𝜽𝑜𝑙𝑑 − 𝜂 ∇𝜽 ℒ ቚ
𝑁 𝜽= 𝜽𝑜𝑙𝑑
𝑖=1

• Batch gradient descent treat the entire training set as a single batch
• Updates the parameter vector after each full pass (epoch) over the entire dataset

11
Stochastic Gradient Descent
• Stochastic gradient descent: we can use a noisy (or stochastic) estimate of the
gradient from a single training example to update the parameter vector 𝜽

𝑁
1
∇𝜽 ℒ = ෍ ∇𝜽 ℓ 𝑖 𝜽𝑛𝑒𝑤 ← 𝜽𝑜𝑙𝑑 − 𝜂 ∇𝜽 ℓ 𝑖 ቚ
𝑁 𝜽= 𝜽𝑜𝑙𝑑
𝑖=1

• The algorithm updates the parameters for every single data point
• Pros: SGD can make significant progress before it has even looked at all the data!

• Cons: It is an approximate (rather stochastic) gradient.

No guarantee that each step will decrease the loss

12
Stochastic Gradient Descent
• We see many fluctuations. Why ? Because we are making greedy decisions

• Each data point is trying to push the parameters in a direction most favorable to
it (without being aware of how the parameter update affects other points)

• A parameter update which is locally favorable to one point may harm other
points (its almost as if the data points are competing with each other)

• There is no guarantee that each local greedy move will reduce the global error

• Can we reduce the oscillations by improving our

stochastic estimates of the gradient (currently
estimated from just 1 data point at a time)?

• Yes, let’s look at mini-batch SGD

13
Mini-batch Gradient Descent
• Compute the gradients on a medium-sized set of training examples, called a
mini-batch

• Note that the algorithm updates the parameters after it sees a batch size 𝐵
number of data points

• The stochastic estimates of gradients here are slightly better and less noisy

• Typical batch sizes are 64, 128, 256

14
Mini-batch Gradient Descent performance

The mini-batch size B is a hyperparameter that needs to be set

• Large batches: converge in fewer parameter updates because each

stochastic gradient is less noisy

• Small batches: perform more parameter updates because each

one requires less computation 15
Things to remember
• 𝑁 is the total number of training examples

• 𝐵 is the mini batch size

• 1 epoch = one pass over the entire data

• 1 iteration = one update step of the parameters

Number of iterations
Algorithm Batch size
in 1 epoch
Batch GD 𝑁 1
SGD 1 𝑁

Mini-batch GD 𝐵 𝑁ൗ
𝐵
PyTorch code mini-batch SGD

17
PyTorch code mini-batch SGD

18
Gradient descent-based optimizers

Problems and workarounds

Local Optima

19
Local optima
• If a function is convex, it has a global minimum and no local minima

ℒ ℒ

𝑤 𝑤

• Convex functions are very convenient for

optimization since starting from any point
in the parameter space, gradient descent
will eventually reach the minimum

• Non-convex functions have multiple local

minima

20
Local optima
• Unfortunately, training a neural network with hidden units is non-convex,
mainly because of permutation symmetries

• Permutation symmetry: We can re-order the weights in a way that gives the
same loss function for the neural network
21
Local optima
• Training a multilayer neural network with hidden units will have multiple
local minima

• Gradient descent starting from two different initializations may converge to

different local optima
23
Local optima
• In general, it is very hard to diagnose if you are stuck in a bad local minimum

• Workaround: One can try to improve the issue by using random restarts

• Random restarts: initialize the training from several random locations, run
the training procedure from each one, and pick whichever result has the
lowest cost

• Random restart is sometimes done in neural net training, but more often we
just ignore the problem

• In practice, the local optima are usually fine, so we think about training in
terms of converging faster to a local minimum, rather than finding the global
minimum
25
Gradient descent-based optimizers

Problems and workarounds

Symmetry

26
Symmetry
• We start our optimization by initializing the values of weights and biases
• Suppose we initialize all the weights and biases of a neural network to all zeros

• All the hidden activations will be identical (indistiguisable features), and all the
weights feeding into a given hidden unit will have zero derivatives
• No learning will occur
• If the initial weights are zero, multiplying them
by any gradient will set the gradients to be
zero. Due to zero gradient, there will be no
change in the weights

Saddle
points
27
Saddle point

𝜕ℒ
• A saddle point has = 𝟎, even though we are not at a minimum
𝜕𝜽
• We are at a location which is minimum with respect to some directions, and
maximum with respect to others
• When would saddle points be a problem?
• If we’re exactly on the saddle point, then we’re stuck
• If we’re slightly to the side, then we can get unstuck

28
Initialization strategies
• Don’t initialize all the weights and biases to zero!

Idea 1: Constant initialization

● Result: For fully connected neural network: identical gradients,

identical neurons. Bad! Because we want to learn different features

Idea 2: Random weights from a standard Gaussian distribution to break symmetry

Popular initialization schemes

• Xavier Init: For Sigmoid/Tanh activation

• Kaiming He Init: ReLU activation

# of neurons in layer
𝑙 −1

29
Gradient descent-based optimizers

Problems and workarounds

Saturated and dead units

30
Saturated and dead units
• A flat region of the loss surface is called a plateau

𝑤2
𝑤1

• Caused by
• Saturated units whose activations are always near the ends of their dynamic
range/possible values
• Dead units whose activations are always close to zero
31
Saturated and dead units
• A flat region of the loss surface is called a plateau

Examples
• Logistic
ℒ activation

𝑤2
𝑤1
• ReLU
activation

• Caused by
• Saturated units whose activations are always near the ends of their dynamic
range/possible values
• Dead units whose activations are always zero (or close to zero)

• Problem: Gradient signal is zero, cannot update the weights 32

Saturated and dead units: Workaround
• Choose the scale of the random initialization of the weights so that the pre-activations
are in the middle of their domain, e.g. Xavier initialization for sigmoid activations

• ReLU units don’t saturate for positive 𝑧, which is convenient. Unfortunately, they can
die if 𝑧 is consistently negative, so it helps to initialize the biases to a small positive
value (such as 0.1)

33
Gradient descent-based optimizers

Problems and workarounds

Ill-conditioning

34
Ill-conditioning
• Suppose that we have the following dataset for linear regression
𝑦 = 𝑤1 𝑥1 + 𝑤2 𝑥2 + 𝑏

𝑤2
𝑥1 𝑥2 𝑡
0.0023 1002.3 5.12
0.0025 1005.2 4.25
0.0016 988.7 3.45
0.0013 945.9 3.13
⋮ ⋮ ⋮
𝑤1

• What kind of values will 𝑤1 and 𝑤2 take?

• 𝑤1 will take larger values, whereas 𝑤2 will take very small values

35
Ill-conditioning
• Which weight, 𝑤1 or 𝑤2 , will receive a larger gradient descent update?
• 𝑤2 receives a much larger update due to steeper gradient

• 𝑤1 receives a much smaller update due to smaller gradient

𝑤2

𝑤1
• But, here we want to take smaller steps in the steeper direction and larger steps in the
flatter direction
36
Ill-conditioning: Workarounds
• We want larger gradient descent updates in the direction of 𝑤1 and smaller in
the direction of 𝑤2
𝑤2
• One workaround: Normalization

• Center the inputs to zero mean and

unit variance
𝑥𝑗 − 𝜇𝑗 𝑤1
𝑥෤𝑗 =
𝜎𝑗
𝑤2

𝑤1 37
Ill-conditioning: Workarounds
• Not just inputs, but inputs to several hidden layers need to be centered too

• Hidden units may also have non-centered activations, and those are even
harder to deal with
• One trick: replace logistic units (which range from 0 to 1) with tanh units (which
range from -1 to 1)

• Another trick: Batch normalization

• It centers each hidden activation and can speed up training by 1.5-2x
• It normalizes the activations of each layer to unit-variance Gaussians
• Is applied immediately after fully connected/conv layers and before non-
linear activations
38
Batch normalization example

𝑦1 𝑦10
Output layer

(2) (2) (2)

Hidden layers
ℎ1 ℎ2 ℎ32

(1) (1) (1)

ℎ1 ℎ2 ℎ64

𝑥1 𝑥2 𝑥3072 Input layer

https://www.machinecurve.com/index.php/2021/03/29/bat
ch-normalization-with-pytorch/ 39
Recap of what we have seen

Problem Diagnostics Workarounds

Local optima Hard to diagnose Random restarts

Symmetries Check 𝐖, 𝒃 Initialize properly

Dead/Saturated units Plot activation histograms Initial scale of W; ReLU

Batch normalization,
Ill-conditioning (hard)
momentum, ADAM

40