Natural Language Processing
with Deep Learning
CS224N/Ling284
Christopher Manning
Lecture 3: Neural net learning: Gradients by hand (matrix calculus)
and algorithmically (the backpropagation algorithm)
1. Introduction
Assignment 2 is all about making sure you really understand the math of neural networks
… then we’ll let the software do it!
We’ll go through it all quickly today, but this is the one week of quarter to most work
through the readings!!!
This will be a tough week for some! à Make sure to get help if you need it:
Visit office hours! Read tutorial materials on the syllabus!
Thursday will be mainly linguistics! Some people find that tough too. 😉
PyTorch tutorial: 3:30pm Friday in Gates B01
A great chance to get an intro to PyTorch, a key deep learning package, before Ass 3!
2
NER: Binary classification for center word being location
• We do supervised training and want high score if it’s a location
1
𝐽! 𝜃 = 𝜎 𝑠 =
1 + 𝑒 "#
predicted model
probability of class
f = Some element-
wise non-linear
function, e.g.,
logistic, tanh, ReLU
∈ R5d
x = [ xmuseums xin xParis xare xamazing ] Embedding of
1-hot words
3
7. Neural computation
Original McCulloch & Pitts
1943 threshold unit:
𝟏(𝑊𝑥 > 𝜃)
= 𝟏(𝑊𝑥 − 𝜃 > 0)
This function has no slope,
so, no gradient-based learning
4
Non-linearities, old and new
(Rectified Linear Unit) Leaky ReLU /
logistic (“sigmoid”) tanh hard tanh ReLU Parametric ReLU
ReLU 𝑧 = max(𝑧, 0)
1 1
0
0
0 −1
tanh is just a rescaled and shifted sigmoid (2×as steep, [−1,1]): GELU arXiv:1606.08415
Swish arXiv:1710.05941
tanh(z) = 2logistic(2z) −1 GELU 𝑥
swish 𝑥 = 𝑥 ' logistic(𝑥)
= 𝑥 1 𝑃 𝑋 ≤ 𝑥 , 𝑋~𝑁(0,1)
Logistic and tanh are still used (e.g., logistic to get a probability) ≈ 𝑥 1 logistic(1.702𝑥)
However, now, for deep networks, the first thing to try is ReLU: it
trains quickly and performs well due to good gradient backflow.
ReLU has a negative “dead zone” that recent proposals mitigate 0
GELU is frequently used with Transformers (BERT, RoBERTa, etc.)
Non-linearities (i.e., “f ” on previous slide): Why they’re needed
• Neural networks do function approximation,
e.g., regression or classification
• Without non-linearities, deep neural networks
can’t do anything more than a linear transform
• Extra layers could just be compiled down into a
single linear transform: W1 W2 x = Wx
• But, with more layers that include non-linearities,
they can approximate any complex function!
6
Training with “cross entropy loss” – you use this in PyTorch!
• Until now, our objective was stated as to maximize the probability of the correct class y
or equivalently we can minimize the negative log probability of that class
• Now restated in terms of cross entropy, a concept from information theory
• Let the true probability distribution be p; let our computed model probability be q
• The cross entropy is:
• Assuming a ground truth (or true or gold or target) probability distribution that is 1 at
the right class and 0 everywhere else, p = [0, …, 0, 1, 0, …, 0], then:
• Because of one-hot p, the only term left is the negative log probability of the true
class yi: − log 𝑝(𝑦. |𝑥. )
Cross entropy can be used in other ways with a more interesting p,
but for now just know that you’ll want to use it as the loss in PyTorch
7
Remember: Stochastic Gradient Descent
Update equation:
𝛼 = step size or learning rate
67 8
i.e., for each parameter: 𝜃/012 = 𝜃/345 −𝛼
68!"#$
In deep learning, 𝜃 includes the data representation (e.g., word vectors) too!
How can we compute ∇8 𝐽(𝜃)?
1. By hand
2. Algorithmically: the backpropagation algorithm
8
Lecture Plan
Lecture 4: Gradients by hand and algorithmically
1. Introduction (10 mins)
2. Matrix calculus (35 mins)
3. Backpropagation (35 mins)
9
Computing Gradients by Hand
• Matrix calculus: Fully vectorized gradients
• “Multivariable calculus is just like single-variable calculus if you use matrices”
• Much faster and more useful than non-vectorized gradients
• But doing a non-vectorized gradient can be good for intuition; recall the first
lecture for an example
• Lecture notes and matrix calculus notes cover this material in more detail
• You might also review Math 51, which has an online textbook:
http://web.stanford.edu/class/math51/textbook.html
10
Gradients
• Given a function with 1 output and 1 input
𝑓 𝑥 = 𝑥9
• It’s gradient (slope) is its derivative
5:
5;
= 3𝑥 <
“How much will the output change if we change the input a bit?”
At x = 1 it changes about 3 times as much: 1.013 = 1.03
At x = 4 it changes about 48 times as much: 4.013 = 64.48
11
Gradients
• Given a function with 1 output and n inputs
• Its gradient is a vector of partial derivatives with
respect to each input
12
Jacobian Matrix: Generalization of the Gradient
• Given a function with m outputs and n inputs
• It’s Jacobian is an m x n matrix of partial derivatives
13
Chain Rule
• For composition of one-variable functions: multiply derivatives
• For multiple variables functions: multiply Jacobians
14
Example Jacobian: Elementwise activation Function
15
Example Jacobian: Elementwise activation Function
Function has n outputs and n inputs → n by n Jacobian
16
Example Jacobian: Elementwise activation Function
17
Example Jacobian: Elementwise activation Function
18
Example Jacobian: Elementwise activation Function
19
Other Jacobians
• Compute these at home for practice!
• Check your answers with the lecture notes
20
Other Jacobians
• Compute these at home for practice!
• Check your answers with the lecture notes
21
Other Jacobians
Fine print: This is the correct Jacobian.
Later we discuss the “shape convention”;
using it the answer would be h.
• Compute these at home for practice!
• Check your answers with the lecture notes
22
Other Jacobians
• Compute these at home for practice!
• Check your answers with the lecture notes
23
Back to our Neural Net!
x = [ xmuseums xin xParis xare xamazing ]
24
Back to our Neural Net!
• Let’s find
• Really, we care about the gradient of the loss Jt but we
will compute the gradient of the score for simplicity
x = [ xmuseums xin xParis xare xamazing ]
25
1. Break up equations into simple pieces
Carefully define your variables and keep track of their dimensionality!
26
2. Apply the chain rule
27
2. Apply the chain rule
28
2. Apply the chain rule
29
2. Apply the chain rule
30
3. Write out the Jacobians
Useful Jacobians from previous slide
31
3. Write out the Jacobians
𝒖!
Useful Jacobians from previous slide
32
3. Write out the Jacobians
𝒖!
Useful Jacobians from previous slide
33
3. Write out the Jacobians
𝒖!
Useful Jacobians from previous slide
34
3. Write out the Jacobians
𝒖!
𝒖! .
Useful Jacobians from previous slide
⊙ = Hadamard product =
element-wise multiplication
of 2 vectors to give vector
35
Re-using Computation
• Suppose we now want to compute
• Using the chain rule again:
36
Re-using Computation
• Suppose we now want to compute
• Using the chain rule again:
The same! Let’s avoid duplicated computation …
37
Re-using Computation
• Suppose we now want to compute
• Using the chain rule again:
𝒖!
𝛿 is the upstream gradient (“error signal”)
38
Derivative with respect to Matrix: Output shape
• What does look like?
• 1 output, nm inputs: 1 by nm Jacobian?
• Inconvenient to then do
39
Derivative with respect to Matrix: Output shape
• What does look like?
• 1 output, nm inputs: 1 by nm Jacobian?
• Inconvenient to then do
• Instead, we leave pure math and use the shape convention:
the shape of the gradient is the shape of the parameters!
• So is n by m:
40
Derivative with respect to Matrix
• What is
• is going to be in our answer
• The other term should be because
• Answer is:
𝛿 is upstream gradient (“error signal”) at 𝑧
𝑥 is local input signal
41
Why the Transposes?
• Hacky answer: this makes the dimensions work out!
• Useful trick for checking your work!
• Full explanation in the lecture notes
• Each input goes to each output – you want to get outer product
42
Deriving local input gradient in backprop
"𝒛
• For "𝑾 in our equation:
𝜕𝑠 𝜕𝒛 𝜕
=𝜹 =𝜹 (𝑾𝒙 + 𝒃)
𝜕𝑾 𝜕𝑾 𝜕𝑾
• Let’s consider the derivative of a single weight Wij
• Wij only contributes to zi u2
• For example: W23 is only
s
used to compute z2 not z1 f(z1)= h1 h2 =f(z2)
W23
𝜕𝑧. 𝜕
= 𝑾.> 𝒙 + 𝑏. b2
𝜕𝑊./ 𝜕𝑊./
6
= ∑5@AB 𝑊.@ 𝑥@ = 𝑥/ x1 x2 x3 +1
6?%!
43
What shape should derivatives be?
• Similarly, is a row vector
• But shape convention says our gradient should be a column vector because b is
a column vector …
• Disagreement between Jacobian form (which makes the chain rule
easy) and the shape convention (which makes implementing SGD easy)
• We expect answers in the assignment to follow the shape convention
• But Jacobian form is useful for computing the answers
44
What shape should derivatives be?
Two options for working through specific problems:
1. Use Jacobian form as much as possible, reshape to
follow the shape convention at the end:
• What we just did. But at the end transpose to make the
derivative a column vector, resulting in
2. Always follow the shape convention
• Look at dimensions to figure out when to transpose and/or
reorder terms
• The error message 𝜹 that arrives at a hidden layer has the
same dimensionality as that hidden layer
45
3. Backpropagation
We’ve almost shown you backpropagation
It’s taking derivatives and using the (generalized, multivariate, or matrix)
chain rule
Other trick:
We re-use derivatives computed for higher layers in computing
derivatives for lower layers to minimize computation
46
Computation Graphs and Backpropagation
• Software represents our neural
net equations as a graph
• Source nodes: inputs
• Interior nodes: operations
+
47
Computation Graphs and Backpropagation
• Software represents our neural
net equations as a graph
• Source nodes: inputs
• Interior nodes: operations
• Edges pass along result of the
operation
+
48
Computation Graphs and Backpropagation
• Software represents our neural
net equations as a graph
• Source nodes: inputs
• “Forward Propagation”
Interior nodes: operations
• Edges pass along result of the
operation
+
49
Backpropagation
• Then go backwards along edges
• Pass along gradients
+
50
Backpropagation: Single Node
• Node receives an “upstream gradient”
• Goal is to pass on the correct
“downstream gradient”
Downstream Upstream
51 gradient gradient
Backpropagation: Single Node
• Each node has a local gradient
• The gradient of its output with
respect to its input
Downstream Local Upstream
52 gradient gradient gradient
Backpropagation: Single Node
• Each node has a local gradient
• The gradient of its output with
respect to its input
Chain
rule!
Downstream Local Upstream
53 gradient gradient gradient
Backpropagation: Single Node
• Each node has a local gradient
• The gradient of its output with
respect to its input
• [downstream gradient] = [upstream gradient] x [local gradient]
Downstream Local Upstream
54 gradient gradient gradient
Backpropagation: Single Node
• What about nodes with multiple inputs?
55
Backpropagation: Single Node
• Multiple inputs → multiple local gradients
Downstream Local Upstream
gradients gradients gradient
56
An Example
57
An Example
Forward prop steps
*
max
58
An Example
Forward prop steps
2
+ 3
6
2 *
2
max
0
59
An Example
Forward prop steps Local gradients
2
+ 3
6
2 *
2
max
0
60
An Example
Forward prop steps Local gradients
2
+ 3
6
2 *
2
max
0
61
An Example
Forward prop steps Local gradients
2
+ 3
6
2 *
2
max
0
62
An Example
Forward prop steps Local gradients
2
+ 3
6
2 *
2
max
0
63
An Example
Forward prop steps Local gradients
2
+ 3
1*2 = 2 6
2 * 1
2
max 1*3 = 3
0
upstream * local = downstream
64
An Example
Forward prop steps Local gradients
2
+ 3
2
6
2 * 1
2
3*1 = 3
max 3
0
3*0 = 0 upstream * local = downstream
65
An Example
Forward prop steps Local gradients
1
2*1 = 2
2
+ 3
2
2*1 = 2 6
2 * 1
2
3
max 3
0
0 upstream * local = downstream
66
An Example
Forward prop steps Local gradients
1
2
2
+ 3
2
2 6
2 * 1
2
3
max 3
0
0
67
Gradients sum at outward branches
68
Gradients sum at outward branches
69
Node Intuitions
• + “distributes” the upstream gradient to each summand
1
2
2
+ 3
2
2 6
2 * 1
2
max
0
70
Node Intuitions
• + “distributes” the upstream gradient to each summand
• max “routes” the upstream gradient
2
+ 3
6
2 * 1
2
3
max 3
0
0
71
Node Intuitions
• + “distributes” the upstream gradient
• max “routes” the upstream gradient
• * “switches” the upstream gradient
2
+ 3
2
6
2 * 1
2
max 3
0
72
Efficiency: compute all gradients at once
• Incorrect way of doing backprop:
• First compute
* +
73
Efficiency: compute all gradients at once
• Incorrect way of doing backprop:
• First compute
• Then independently compute
• Duplicated computation!
* +
74
Efficiency: compute all gradients at once
• Correct way:
• Compute all the gradients at once
• Analogous to using 𝜹 when we
computed gradients by hand
* +
75
Back-Prop in General Computation Graph
1. Fprop: visit nodes in topological sort order
Single scalar output - Compute value of node given predecessors
2. Bprop:
- initialize output gradient = 1
… - visit nodes in reverse order:
Compute gradient wrt each node using
gradient wrt successors
… = successors of
Done correctly, big O() complexity of fprop and
bprop is the same
In general, our nets have regular layer-structure
Inputs and so we can use matrices and Jacobians…
76
Automatic Differentiation
• The gradient computation can be
automatically inferred from the symbolic
expression of the fprop
• Each node type needs to know how to
compute its output and how to compute
the gradient wrt its inputs given the
gradient wrt its output
• Modern DL frameworks (Tensorflow,
PyTorch, etc.) do backpropagation for
you but mainly leave layer/node writer
to hand-calculate the local derivative
77
Backprop Implementations
78
Implementation: forward/backward API
79
Implementation: forward/backward API
80
Manual Gradient checking: Numeric Gradient
• For small h (≈ 1e-4),
• Easy to implement correctly
• But approximate and very slow:
• You have to recompute f for every parameter of our model
• Useful for checking your implementation
• In the old days, we hand-wrote everything, doing this everywhere was the key test
• Now much less needed; you can use it to check layers are correctly implemented
81
Summary
We’ve mastered the core technology of neural nets! 🎉 🎉 🎉
• Backpropagation: recursively (and hence efficiently) apply the chain rule
along computation graph
• [downstream gradient] = [upstream gradient] x [local gradient]
• Forward pass: compute results of operations and save intermediate
values
• Backward pass: apply chain rule to compute gradients
82
Why learn all these details about gradients?
• Modern deep learning frameworks compute gradients for you!
• Come to the PyTorch introduction this Friday!
• But why take a class on compilers or systems when they are implemented for you?
• Understanding what is going on under the hood is useful!
• Backpropagation doesn’t always work perfectly out of the box
• Understanding why is crucial for debugging and improving models
• See Karpathy article (in syllabus):
• https://medium.com/@karpathy/yes-you-should-understand-backprop-e2f06eab496b
• Example in future lecture: exploding and vanishing gradients
83
