Neural Networks and Neural Language Models

Speech and Language Processing. Daniel Jurafsky & James H. Martin. Copyright © 2023.
All
rights reserved. Draft of January 7, 2023.
CHAPTER
Neural Networks and Neural

7 Language Models
“[M]achines of this character can behave in a very complicated manner when

the number of units is large.”
Alan Turing (1948) “Intelligent Machines”, page 6
Neural networks are a fundamental computational tool for language process-

ing, and a very old one. They are called neural because their origins lie in the
McCulloch-Pitts neuron (McCulloch and Pitts, 1943), a simplified model of the
human neuron as a kind of computing element that could be described in terms of
propositional logic. But the modern use in language processing no longer draws on
these early biological inspirations.
Instead, a modern neural network is a network of small computing units, each
of which takes a vector of input values and produces a single output value. In this
chapter we introduce the neural net applied to classification. The architecture we
feedforward introduce is called a feedforward network because the computation proceeds iter-
atively from one layer of units to the next. The use of modern neural nets is often
deep learning called deep learning, because modern networks are often deep (have many layers).
Neural networks share much of the same mathematics as logistic regression. But
neural networks are a more powerful classifier than logistic regression, and indeed a
minimal neural network (technically one with a single ‘hidden layer’) can be shown
to learn any function.
Neural net classifiers are different from logistic regression in another way. With
logistic regression, we applied the regression classifier to many different tasks by
developing many rich kinds of feature templates based on domain knowledge. When
working with neural networks, it is more common to avoid most uses of rich hand-
derived features, instead building neural networks that take raw words as inputs
and learn to induce features as part of the process of learning to classify. We saw
examples of this kind of representation learning for embeddings in Chapter 6. Nets
that are very deep are particularly good at representation learning. For that reason
deep neural nets are the right tool for tasks that offer sufficient data to learn features
automatically.
In this chapter we’ll introduce feedforward networks as classifiers, and also ap-
ply them to the simple task of language modeling: assigning probabilities to word
sequences and predicting upcoming words. In subsequent chapters we’ll introduce
many other aspects of neural models, such as recurrent neural networks (Chap-
ter 9), the Transformer (Chapter 10), and masked language modeling (Chapter 11).
2 C HAPTER 7 • N EURAL N ETWORKS AND N EURAL L ANGUAGE M ODELS
7.1 Units
The building block of a neural network is a single computational unit. A unit takes
a set of real valued numbers as input, performs some computation on them, and
produces an output.
At its heart, a neural unit is taking a weighted sum of its inputs, with one addi-
bias term tional term in the sum called a bias term. Given a set of inputs x1 ...xn , a unit has
a set of corresponding weights w1 ...wn and a bias b, so the weighted sum z can be
represented as: X
z = b+ wi xi (7.1)
i
Often it’s more convenient to express this weighted sum using vector notation; recall
vector from linear algebra that a vector is, at heart, just a list or array of numbers. Thus
we’ll talk about z in terms of a weight vector w, a scalar bias b, and an input vector
x, and we’ll replace the sum with the convenient dot product:
z = w·x+b (7.2)
As defined in Eq. 7.2, z is just a real valued number.
Finally, instead of using z, a linear function of x, as the output, neural units
apply a non-linear function f to z. We will refer to the output of this function as
activation the activation value for the unit, a. Since we are just modeling a single unit, the
activation for the node is in fact the final output of the network, which we’ll generally
call y. So the value y is defined as:
y = a = f (z)
We’ll discuss three popular non-linear functions f () below (the sigmoid, the tanh,
and the rectified linear unit or ReLU) but it’s pedagogically convenient to start with
sigmoid the sigmoid function since we saw it in Chapter 5:
1
y = σ (z) = (7.3)
1 + e−z
The sigmoid (shown in Fig. 7.1) has a number of advantages; it maps the output
into the range (0, 1), which is useful in squashing outliers toward 0 or 1. And it’s
differentiable, which as we saw in Section ?? will be handy for learning.
Figure 7.1 The sigmoid function takes a real value and maps it to the range (0, 1). It is
nearly linear around 0 but outlier values get squashed toward 0 or 1.
Substituting Eq. 7.2 into Eq. 7.3 gives us the output of a neural unit:
1
y = σ (w · x + b) = (7.4)
1 + exp(−(w · x + b))
7.1 • U NITS 3
Fig. 7.2 shows a final schematic of a basic neural unit. In this example the unit
takes 3 input values x1 , x2 , and x3 , and computes a weighted sum, multiplying each
value by a weight (w1 , w2 , and w3 , respectively), adds them to a bias term b, and then
passes the resulting sum through a sigmoid function to result in a number between 0
and 1.
x1 w1
w2 z a
x2 ∑ σ y
w3
x3 b
+1
Figure 7.2 A neural unit, taking 3 inputs x1 , x2 , and x3 (and a bias b that we represent as a
weight for an input clamped at +1) and producing an output y. We include some convenient
intermediate variables: the output of the summation, z, and the output of the sigmoid, a. In
this case the output of the unit y is the same as a, but in deeper networks we’ll reserve y to
mean the final output of the entire network, leaving a as the activation of an individual node.
Let’s walk through an example just to get an intuition. Let’s suppose we have a
unit with the following weight vector and bias:
w = [0.2, 0.3, 0.9]

b = 0.5
What would this unit do with the following input vector:
x = [0.5, 0.6, 0.1]
The resulting output y would be:

1 1 1
y = σ (w · x + b) = = = = .70
1 + e−(w·x+b) 1 + e−(.5∗.2+.6∗.3+.1∗.9+.5) 1 + e−0.87
In practice, the sigmoid is not commonly used as an activation function. A function
tanh that is very similar but almost always better is the tanh function shown in Fig. 7.3a;
tanh is a variant of the sigmoid that ranges from -1 to +1:
ez − e−z
y = tanh(z) = (7.5)
ez + e−z
The simplest activation function, and perhaps the most commonly used, is the rec-
ReLU tified linear unit, also called the ReLU, shown in Fig. 7.3b. It’s just the same as z
when z is positive, and 0 otherwise:
y = ReLU(z) = max(z, 0) (7.6)
These activation functions have different properties that make them useful for differ-
ent language applications or network architectures. For example, the tanh function
has the nice properties of being smoothly differentiable and mapping outlier values
toward the mean. The rectifier function, on the other hand, has nice properties that
(a) (b)
Figure 7.3 The tanh and ReLU activation functions.
result from it being very close to linear. In the sigmoid or tanh functions, very high
saturated values of z result in values of y that are saturated, i.e., extremely close to 1, and have
derivatives very close to 0. Zero derivatives cause problems for learning, because as
we’ll see in Section 7.6, we’ll train networks by propagating an error signal back-
wards, multiplying gradients (partial derivatives) from each layer of the network;
gradients that are almost 0 cause the error signal to get smaller and smaller until it is
vanishing
gradient too small to be used for training, a problem called the vanishing gradient problem.
Rectifiers don’t have this problem, since the derivative of ReLU for high values of z
is 1 rather than very close to 0.
7.2 The XOR problem

Early in the history of neural networks it was realized that the power of neural net-
works, as with the real neurons that inspired them, comes from combining these
units into larger networks.
One of the most clever demonstrations of the need for multi-layer networks was
the proof by Minsky and Papert (1969) that a single neural unit cannot compute
some very simple functions of its input. Consider the task of computing elementary
logical functions of two inputs, like AND, OR, and XOR. As a reminder, here are
the truth tables for those functions:
AND OR XOR
x1 x2 y x1 x2 y x1 x2 y
0 0 0 0 0 0 0 0 0
0 1 0 0 1 1 0 1 1
1 0 0 1 0 1 1 0 1
1 1 1 1 1 1 1 1 0
perceptron This example was first shown for the perceptron, which is a very simple neural
unit that has a binary output and does not have a non-linear activation function. The
output y of a perceptron is 0 or 1, and is computed as follows (using the same weight
w, input x, and bias b as in Eq. 7.2):

0, if w · x + b ≤ 0
y= (7.7)
1, if w · x + b > 0
7.2 • T HE XOR PROBLEM 5
It’s very easy to build a perceptron that can compute the logical AND and OR
functions of its binary inputs; Fig. 7.4 shows the necessary weights.
x1 x1
1 1
x2 1 x2 1
-1 0
+1 +1
(a) (b)
Figure 7.4 The weights w and bias b for perceptrons for computing logical functions. The
inputs are shown as x1 and x2 and the bias as a special node with value +1 which is multiplied
with the bias weight b. (a) logical AND, with weights w1 = 1 and w2 = 1 and bias weight
b = −1. (b) logical OR, with weights w1 = 1 and w2 = 1 and bias weight b = 0. These
weights/biases are just one from an infinite number of possible sets of weights and biases that
would implement the functions.
It turns out, however, that it’s not possible to build a perceptron to compute
logical XOR! (It’s worth spending a moment to give it a try!)
The intuition behind this important result relies on understanding that a percep-
tron is a linear classifier. For a two-dimensional input x1 and x2 , the perceptron
equation, w1 x1 + w2 x2 + b = 0 is the equation of a line. (We can see this by putting
it in the standard linear format: x2 = (−w1 /w2 )x1 + (−b/w2 ).) This line acts as a
decision
boundary decision boundary in two-dimensional space in which the output 0 is assigned to all
inputs lying on one side of the line, and the output 1 to all input points lying on the
other side of the line. If we had more than 2 inputs, the decision boundary becomes
a hyperplane instead of a line, but the idea is the same, separating the space into two
categories.
Fig. 7.5 shows the possible logical inputs (00, 01, 10, and 11) and the line drawn
by one possible set of parameters for an AND and an OR classifier. Notice that there
is simply no way to draw a line that separates the positive cases of XOR (01 and 10)
linearly
separable from the negative cases (00 and 11). We say that XOR is not a linearly separable
function. Of course we could draw a boundary with a curve, or some other function,
but not a single line.
7.2.1 The solution: neural networks

While the XOR function cannot be calculated by a single perceptron, it can be cal-
culated by a layered network of perceptron units. Rather than see this with networks
of simple perceptrons, however, let’s see how to compute XOR using two layers of
ReLU-based units following Goodfellow et al. (2016). Fig. 7.6 shows a figure with
the input being processed by two layers of neural units. The middle layer (called
h) has two units, and the output layer (called y) has one unit. A set of weights and
biases are shown for each ReLU that correctly computes the XOR function.
Let’s walk through what happens with the input x = [0, 0]. If we multiply each
input value by the appropriate weight, sum, and then add the bias b, we get the vector
[0, -1], and we then apply the rectified linear transformation to give the output of the
h layer as [0, 0]. Now we once again multiply by the weights, sum, and add the
bias (0 in this case) resulting in the value 0. The reader should work through the
computation of the remaining 3 possible input pairs to see that the resulting y values
are 1 for the inputs [0, 1] and [1, 0] and 0 for [0, 0] and [1, 1].
x2 x2 x2
1 1 1
?
0 x1 0 x1 0 x1
0 1 0 1 0 1
a) x1 AND x2 b) x1 OR x2 c) x1 XOR x2
Figure 7.5 The functions AND, OR, and XOR, represented with input x1 on the x-axis and input x2 on the
y-axis. Filled circles represent perceptron outputs of 1, and white circles perceptron outputs of 0. There is no
way to draw a line that correctly separates the two categories for XOR. Figure styled after Russell and Norvig
(2002).
x1 1 h1
1
1
y1
1
-2
x2 1 h2 0
0
-1
+1 +1
Figure 7.6 XOR solution after Goodfellow et al. (2016). There are three ReLU units, in
two layers; we’ve called them h1 , h2 (h for “hidden layer”) and y1 . As before, the numbers
on the arrows represent the weights w for each unit, and we represent the bias b as a weight
on a unit clamped to +1, with the bias weights/units in gray.
It’s also instructive to look at the intermediate results, the outputs of the two
hidden nodes h1 and h2 . We showed in the previous paragraph that the h vector for
the inputs x = [0, 0] was [0, 0]. Fig. 7.7b shows the values of the h layer for all
4 inputs. Notice that hidden representations of the two input points x = [0, 1] and
x = [1, 0] (the two cases with XOR output = 1) are merged to the single point h =
[1, 0]. The merger makes it easy to linearly separate the positive and negative cases
of XOR. In other words, we can view the hidden layer of the network as forming a
representation of the input.
In this example we just stipulated the weights in Fig. 7.6. But for real examples
the weights for neural networks are learned automatically using the error backprop-
agation algorithm to be introduced in Section 7.6. That means the hidden layers will
learn to form useful representations. This intuition, that neural networks can auto-
matically learn useful representations of the input, is one of their key advantages,
and one that we will return to again and again in later chapters.
7.3 • F EEDFORWARD N EURAL N ETWORKS 7
x2 h2
1 1
0 x1 0
h1
0 1 0 1 2
a) The original x space b) The new (linearly separable) h space

Figure 7.7 The hidden layer forming a new representation of the input. (b) shows the
representation of the hidden layer, h, compared to the original input representation x in (a).
Notice that the input point [0, 1] has been collapsed with the input point [1, 0], making it
possible to linearly separate the positive and negative cases of XOR. After Goodfellow et al.
(2016).
7.3 Feedforward Neural Networks

Let’s now walk through a slightly more formal presentation of the simplest kind of
feedforward neural network, the feedforward network. A feedforward network is a multilayer
network
network in which the units are connected with no cycles; the outputs from units in
each layer are passed to units in the next higher layer, and no outputs are passed
back to lower layers. (In Chapter 9 we’ll introduce networks with cycles, called
recurrent neural networks.)
For historical reasons multilayer networks, especially feedforward networks, are
multi-layer
perceptrons sometimes called multi-layer perceptrons (or MLPs); this is a technical misnomer,
MLP since the units in modern multilayer networks aren’t perceptrons (perceptrons are
purely linear, but modern networks are made up of units with non-linearities like
sigmoids), but at some point the name stuck.
Simple feedforward networks have three kinds of nodes: input units, hidden
units, and output units.
Fig. 7.8 shows a picture. The input layer x is a vector of simple scalar values just
as we saw in Fig. 7.2.
hidden layer The core of the neural network is the hidden layer h formed of hidden units hi ,
each of which is a neural unit as described in Section 7.1, taking a weighted sum of
its inputs and then applying a non-linearity. In the standard architecture, each layer
fully-connected is fully-connected, meaning that each unit in each layer takes as input the outputs
from all the units in the previous layer, and there is a link between every pair of units
from two adjacent layers. Thus each hidden unit sums over all the input units.
Recall that a single hidden unit has as parameters a weight vector and a bias. We
represent the parameters for the entire hidden layer by combining the weight vector
and bias for each unit i into a single weight matrix W and a single bias vector b for
the whole layer (see Fig. 7.8). Each element W ji of the weight matrix W represents
the weight of the connection from the ith input unit xi to the jth hidden unit h j .
The advantage of using a single matrix W for the weights of the entire layer is
that now the hidden layer computation for a feedforward network can be done very
efficiently with simple matrix operations. In fact, the computation only has three
x1 W U
y1
h1
x2 h2 y2
h3
…
…
xn
0 hn
1
b yn
2
+1
input layer hidden layer output layer
Figure 7.8 A simple 2-layer feedforward network, with one hidden layer, one output layer,
and one input layer (the input layer is usually not counted when enumerating layers).
steps: multiplying the weight matrix by the input vector x, adding the bias vector b,
and applying the activation function g (such as the sigmoid, tanh, or ReLU activation
function defined above).
The output of the hidden layer, the vector h, is thus the following (for this exam-
ple we’ll use the sigmoid function σ as our activation function):
h = σ (Wx + b) (7.8)
Notice that we’re applying the σ function here to a vector, while in Eq. 7.3 it was
applied to a scalar. We’re thus allowing σ (·), and indeed any activation function
g(·), to apply to a vector element-wise, so g[z1 , z2 , z3 ] = [g(z1 ), g(z2 ), g(z3 )].
Let’s introduce some constants to represent the dimensionalities of these vectors
and matrices. We’ll refer to the input layer as layer 0 of the network, and have n0
represent the number of inputs, so x is a vector of real numbers of dimension n0 ,
or more formally x ∈ Rn0 , a column vector of dimensionality [n0 , 1]. Let’s call the
hidden layer layer 1 and the output layer layer 2. The hidden layer has dimensional-
ity n1 , so h ∈ Rn1 and also b ∈ Rn1 (since each hidden unit can take a different bias
value). And the weight matrix W has dimensionality W ∈ Rn1 ×n0 , i.e. [n1 , n0 ].
Take a moment to convince yourselfPn0 that the matrix multiplication in Eq. 7.8 will
compute the value of each h j as σ i=1 W ji xi + b j .
As we saw in Section 7.2, the resulting value h (for hidden but also for hypoth-
esis) forms a representation of the input. The role of the output layer is to take
this new representation h and compute a final output. This output could be a real-
valued number, but in many cases the goal of the network is to make some sort of
classification decision, and so we will focus on the case of classification.
If we are doing a binary task like sentiment classification, we might have a sin-
gle output node, and its scalar value y is the probability of positive versus negative
sentiment. If we are doing multinomial classification, such as assigning a part-of-
speech tag, we might have one output node for each potential part-of-speech, whose
output value is the probability of that part-of-speech, and the values of all the output
nodes must sum to one. The output layer is thus a vector y that gives a probability
distribution across the output nodes.
Let’s see how this happens. Like the hidden layer, the output layer has a weight
matrix (let’s call it U), but some models don’t include a bias vector b in the output
7.3 • F EEDFORWARD N EURAL N ETWORKS 9
layer, so we’ll simplify by eliminating the bias vector in this example. The weight
matrix is multiplied by its input vector (h) to produce the intermediate output z:
z = Uh
There are n2 output nodes, so z ∈ Rn2 , weight matrix U has dimensionality U ∈
Rn2 ×n1 , and element Ui j is the weight from unit j in the hidden layer to unit i in the
output layer.
However, z can’t be the output of the classifier, since it’s a vector of real-valued
numbers, while what we need for classification is a vector of probabilities. There is
normalizing a convenient function for normalizing a vector of real values, by which we mean
converting it to a vector that encodes a probability distribution (all the numbers lie
softmax between 0 and 1 and sum to 1): the softmax function that we saw on page ?? of
Chapter 5. More generally for any vector z of dimensionality d, the softmax is
defined as:
exp(zi )
softmax(zi ) = Pd 1≤i≤d (7.9)
j=1 exp(z j )
Thus for example given a vector

z = [0.6, 1.1, −1.5, 1.2, 3.2, −1.1], (7.10)
the softmax function will normalize it to a probability distribution (shown rounded):

softmax(z) = [0.055, 0.090, 0.0067, 0.10, 0.74, 0.010] (7.11)
You may recall that we used softmax to create a probability distribution from a
vector of real-valued numbers (computed from summing weights times features) in
the multinomial version of logistic regression in Chapter 5.
That means we can think of a neural network classifier with one hidden layer
as building a vector h which is a hidden layer representation of the input, and then
running standard multinomial logistic regression on the features that the network
develops in h. By contrast, in Chapter 5 the features were mainly designed by hand
via feature templates. So a neural network is like multinomial logistic regression,
but (a) with many layers, since a deep neural network is like layer after layer of lo-
gistic regression classifiers; (b) with those intermediate layers having many possible
activation functions (tanh, ReLU, sigmoid) instead of just sigmoid (although we’ll
continue to use σ for convenience to mean any activation function); (c) rather than
forming the features by feature templates, the prior layers of the network induce the
feature representations themselves.
Here are the final equations for a feedforward network with a single hidden layer,
which takes an input vector x, outputs a probability distribution y, and is parameter-
ized by weight matrices W and U and a bias vector b:
h = σ (Wx + b)
z = Uh
y = softmax(z) (7.12)
And just to remember the shapes of all our variables, x ∈ Rn0 , h ∈ Rn1 , b ∈ Rn1 ,
W ∈ Rn1 ×n0 , U ∈ Rn2 ×n1 , and the output vector y ∈ Rn2 . We’ll call this network a 2-
layer network (we traditionally don’t count the input layer when numbering layers,
but do count the output layer). So by this terminology logistic regression is a 1-layer
network.
7.3.1 More details on feedforward networks

Let’s now set up some notation to make it easier to talk about deeper networks of
depth more than 2. We’ll use superscripts in square brackets to mean layer num-
bers, starting at 0 for the input layer. So W[1] will mean the weight matrix for the
(first) hidden layer, and b[1] will mean the bias vector for the (first) hidden layer. n j
will mean the number of units at layer j. We’ll use g(·) to stand for the activation
function, which will tend to be ReLU or tanh for intermediate layers and softmax
for output layers. We’ll use a[i] to mean the output from layer i, and z[i] to mean the
combination of weights and biases W[i] a[i−1] + b[i] . The 0th layer is for inputs, so
we’ll refer to the inputs x more generally as a[0] .
Thus we can re-represent our 2-layer net from Eq. 7.12 as follows:
z[1] = W[1] a[0] + b[1]

a[1] = g[1] (z[1] )
z[2] = W[2] a[1] + b[2]
a[2] = g[2] (z[2] )
ŷ = a[2] (7.13)
Note that with this notation, the equations for the computation done at each layer are
the same. The algorithm for computing the forward step in an n-layer feedforward
network, given the input vector a[0] is thus simply:
for i in 1,...,n
z[i] = W[i] a[i−1] + b[i]
a[i] = g[i] (z[i] )
ŷ = a[n]
The activation functions g(·) are generally different at the final layer. Thus g[2]
might be softmax for multinomial classification or sigmoid for binary classification,
while ReLU or tanh might be the activation function g(·) at the internal layers.
The need for non-linear activation functions One of the reasons we use non-
linear activation functions for each layer in a neural network is that if we did not, the
resulting network is exactly equivalent to a single-layer network. Let’s see why this
is true. Imagine the first two layers of such a network of purely linear layers:
z[1] = W[1] x + b[1]

z[2] = W[2] z[1] + b[2]
We can rewrite the function that the network is computing as:
z[2] = W[2] z[1] + b[2]

= W[2] (W[1] x + b[1] ) + b[2]
= W[2] W[1] x + W[2] b[1] + b[2]
= W0 x + b0 (7.14)
This generalizes to any number of layers. So without non-linear activation functions,

a multilayer network is just a notational variant of a single layer network with a
different set of weights, and we lose all the representational power of multilayer
networks.
7.4 • F EEDFORWARD NETWORKS FOR NLP: C LASSIFICATION 11
Replacing the bias unit In describing networks, we will often use a slightly sim-
plified notation that represents exactly the same function without referring to an ex-
plicit bias node b. Instead, we add a dummy node a0 to each layer whose value will
[0]
always be 1. Thus layer 0, the input layer, will have a dummy node a0 = 1, layer 1
[1]
will have a0 = 1, and so on. This dummy node still has an associated weight, and
that weight represents the bias value b. For example instead of an equation like
h = σ (Wx + b) (7.15)
we’ll use:
h = σ (Wx) (7.16)
But now instead of our vector x having n0 values: x = x1 , . . . , xn0 , it will have n0 +
1 values, with a new 0th dummy value x0 = 1: x = x0 , . . . , xn0 . And instead of
computing each h j as follows:
n0
!
X
hj = σ Wji xi + b j , (7.17)
i=1
we’ll instead use:

n0
!
X
hj = σ Wji xi , (7.18)
i=0
where the value Wj0 replaces what had been b j . Fig. 7.9 shows a visualization.
W U W U
x1 h1 y1 x0=1
h1 y1
h2
x2 y2 x1 h2 y2
h3
…
x2 h3
…
…
…
…
xn
hn
…
0
1
yn hn
b 2 xn 1 yn
+1 0 2
(a) (b)
Figure 7.9 Replacing the bias node (shown in a) with x0 (b).
We’ll continue showing the bias as b when we go over the learning algorithm
in Section 7.6, but then we’ll switch to this simplified notation without explicit bias
terms for the rest of the book.
7.4 Feedforward networks for NLP: Classification

Let’s see how to apply feedforward networks to NLP tasks! In this section we’ll
look at classification tasks like sentiment analysis; in the next section we’ll introduce
neural language modeling.
Let’s begin with a simple 2-layer sentiment classifier. You might imagine tak-
ing our logistic regression classifier from Chapter 5, which corresponds to a 1-layer
network, and just adding a hidden layer. The input element xi could be scalar fea-
tures like those in Fig. ??, e.g., x1 = count(words ∈ doc), x2 = count(positive lexicon
words ∈ doc), x3 = 1 if “no” ∈ doc, and so on. And the output layer ŷ could have
two nodes (one each for positive and negative), or 3 nodes (positive, negative, neu-
tral), in which case ŷ1 would be the estimated probability of positive sentiment, ŷ2
the probability of negative and ŷ3 the probability of neutral. The resulting equations
would be just what we saw above for a 2-layer network (as always, we’ll continue
to use the σ to stand for any non-linearity, whether sigmoid, ReLU or other).
x = [x1 , x2 , ...xN ] (each xi is a hand-designed feature)
h = σ (Wx + b)
z = Uh
ŷ = softmax(z) (7.19)
Fig. 7.10 shows a sketch of this architecture. As we mentioned earlier, adding this
hidden layer to our logistic regression classifier allows the network to represent the
non-linear interactions between features. This alone might give us a better sentiment
classifier.
h1
dessert wordcount x1
=3
h2
y^1 p(+)
positive lexicon x2
was words = 1 y^ 2 p(-)
h3
y^3
…
great count of “no” x3 p(neut)

=0
hdh
Input words x W h U y
[n⨉1] [dh⨉n] [3⨉dh] [3⨉1]
[dh⨉1]
Input layer Hidden layer Output layer
n=3 features softmax
Figure 7.10 Feedforward network sentiment analysis using traditional hand-built features
of the input text.
Most applications of neural networks for NLP do something different, however.

Instead of using hand-built human-engineered features as the input to our classifier,
we draw on deep learning’s ability to learn features from the data by representing
words as embeddings, like the word2vec or GloVe embeddings we saw in Chapter 6.
There are various ways to represent an input for classification. One simple baseline
pooling is to apply some sort of pooling function to the embeddings of all the words in the
input. For example, for a text with n input words/tokens w1 , ..., wn , we can turn the
n embeddings e(w1 ), ..., e(wn ) (each of dimensionality d) into a single embedding
also of dimensionality d by just summing the embeddings, or by taking their mean
(summing and then dividing by n):
n
1X
xmean = e(wi ) (7.20)
n
i=1
7.4 • F EEDFORWARD NETWORKS FOR NLP: C LASSIFICATION 13
There are many other options, like taking the element-wise max. The element-wise
max of a set of n vectors is a new vector whose kth element is the max of the kth
elements of all the n vectors. Here are the equations for this classifier assuming
mean pooling; the architecture is sketched in Fig. 7.11:
x = mean(e(w1 ), e(w2 ), . . . , e(wn ))

h = σ (Wx + b)
z = Uh
h1
embedding for
dessert “dessert” y^ 1 p(+)
pooling h2
was
embedding for ^y p(-)
+
“was” 2
h3
embedding for ^y p(neut)
…
great “great” 3
hdh
Input words x W h U y
[d⨉1] [dh⨉d] [3⨉dh] [3⨉1]
[dh⨉1]
Input layer Hidden layer Output layer

pooled softmax
embedding
Figure 7.11 Feedforward network sentiment analysis using a pooled embedding of the in-
put words.
While Eq. 7.21 shows how to a classify a single example x, in practice we want
to efficiently classify an entire test set of m examples. We do this by vectoring the
process, just as we saw with logistic regression; instead of using for-loops to go
through each example, we’ll use matrix multiplication to do the entire computation
of an entire test set at once. First, we pack all the input feature vectors for each input
x into a single input matrix X, with each row i a row vector consisting of the pooled
embedding for input example x(i) (i.e., the vector x(i) ). If the dimensionality of our
pooled input embedding is d, X will be a matrix of shape [m × d].
We will then need to slightly modify Eq. 7.21. X is of shape [m × d] and W is of
shape [dh × d], so we’ll have to reorder how we multiply X and W and transpose W
so they correctly multiply to yield a matrix H of shape [m × dh ]. The bias vector b
from Eq. 7.21 of shape [1 × dh ] will now have to be replicated into a matrix of shape
[m × dh ]. We’ll need to similarly reorder the next step and transpose U. Finally, our
output matrix Ŷ will be of shape [m × 3] (or more generally [m × do ], where do is
the number of output classes), with each row i of our output matrix Ŷ consisting of
the output vector ŷ(i) .‘ Here are the final equations for computing the output class
distribution for an entire test set:
H = σ (XW| + b)
Z = HU|
Ŷ = softmax(Z) (7.22)
The idea of using word2vec or GloVe embeddings as our input representation—

and more generally the idea of relying on another algorithm to have already learned
pretraining an embedding representation for our input words—is called pretraining. Using
pretrained embedding representations, whether simple static word embeddings like
word2vec or the much more powerful contextual embeddings we’ll introduce in
Chapter 11, is one of the central ideas of deep learning. (It’s also possible, how-
ever, to train the word embeddings as part of an NLP task; we’ll talk about how to
do this in Section 7.7 in the context of the neural language modeling task.)
7.5 Feedforward Neural Language Modeling

As our second application of feedforward networks, let’s consider language model-
ing: predicting upcoming words from prior word context. Neural language modeling
is an important NLP task in itself, and it plays a role in many important algorithms
for tasks like machine translation, summarization, speech recognition, grammar cor-
rection, and dialogue. We’ll describe simple feedforward neural language models,
first introduced by Bengio et al. (2003). While modern neural language models use
more powerful architectures like the recurrent nets or transformer networks to be
introduced in Chapter 9, the feedforward language model introduces many of the
important concepts of neural language modeling.
Neural language models have many advantages over the n-gram language mod-
els of Chapter 3. Compared to n-gram models, neural language models can handle
much longer histories, can generalize better over contexts of similar words, and are
more accurate at word-prediction. On the other hand, neural net language models
are much more complex, are slower and need more energy to train, and are less in-
terpretable than n-gram models, so for many (especially smaller) tasks an n-gram
language model is still the right tool.
A feedforward neural language model (LM) is a feedforward network that takes
as input at time t a representation of some number of previous words (wt−1 , wt−2 ,
etc.) and outputs a probability distribution over possible next words. Thus—like the
n-gram LM—the feedforward neural LM approximates the probability of a word
given the entire prior context P(wt |w1:t−1 ) by approximating based on the N − 1
previous words:
P(wt |w1 , . . . , wt−1 ) ≈ P(wt |wt−N+1 , . . . , wt−1 ) (7.23)
In the following examples we’ll use a 4-gram example, so we’ll show a neural net to
estimate the probability P(wt = i|wt−3 , wt−2 , wt−1 ).
Neural language models represent words in this prior context by their embed-
dings, rather than just by their word identity as used in n-gram language models.
Using embeddings allows neural language models to generalize better to unseen
data. For example, suppose we’ve seen this sentence in training:
I have to make sure that the cat gets fed.
but have never seen the words “gets fed” after the word “dog”. Our test set has the
prefix “I forgot to make sure that the dog gets”. What’s the next word? An n-gram
language model will predict “fed” after “that the cat gets”, but not after “that the dog
gets”. But a neural LM, knowing that “cat” and “dog” have similar embeddings, will
be able to generalize from the “cat” context to assign a high enough probability to
“fed” even after seeing “dog”.
7.5 • F EEDFORWARD N EURAL L ANGUAGE M ODELING 15
7.5.1 Forward inference in the neural language model

forward Let’s walk through forward inference or decoding for neural language models.
inference
Forward inference is the task, given an input, of running a forward pass on the
network to produce a probability distribution over possible outputs, in this case next
words.
We first represent each of the N previous words as a one-hot vector of length
one-hot vector |V |, i.e., with one dimension for each word in the vocabulary. A one-hot vector is
a vector that has one element equal to 1—in the dimension corresponding to that
word’s index in the vocabulary— while all the other elements are set to zero. Thus
in a one-hot representation for the word “toothpaste”, supposing it is V5 , i.e., index
5 in the vocabulary, x5 = 1, and xi = 0 ∀i 6= 5, as shown here:
[0 0 0 0 1 0 0 ... 0 0 0 0]
1 2 3 4 5 6 7 ... ... |V|
The feedforward neural language model (sketched in Fig. 7.13) has a moving
window that can see N words into the past. We’ll let N equal 3, so the 3 words
wt−1 , wt−2 , and wt−3 are each represented as a one-hot vector. We then multiply
these one-hot vectors by the embedding matrix E. The embedding weight matrix E
has a column for each word, each a column vector of d dimensions, and hence has
dimensionality d × |V |. Multiplying by a one-hot vector that has only one non-zero
element xi = 1 simply selects out the relevant column vector for word i, resulting in
the embedding for word i, as shown in Fig. 7.12.
|V| 1 1
d E ✕ 5 = d
5 |V|
e5
Figure 7.12 Selecting the embedding vector for word V5 by multiplying the embedding
matrix E with a one-hot vector with a 1 in index 5.
The 3 resulting embedding vectors are concatenated to produce e, the embedding

layer. This is followed by a hidden layer and an output layer whose softmax produces
a probability distribution over words. For example y42 , the value of output node 42,
is the probability of the next word wt being V42 , the vocabulary word with index 42
(which is the word ‘fish’ in our example).
Here’s the algorithm in detail for our mini example:
1. Select three embeddings from E: Given the three previous words, we look
up their indices, create 3 one-hot vectors, and then multiply each by the em-
bedding matrix E. Consider wt−3 . The one-hot vector for ‘for’ (index 35) is
multiplied by the embedding matrix E, to give the first part of the first hidden
embedding
layer layer, the embedding layer. Since each column of the input matrix E is an
embedding for a word, and the input is a one-hot column vector xi for word
Vi , the embedding layer for input w will be Exi = ei , the embedding for word
i. We now concatenate the three embeddings for the three context words to
produce the embedding layer e.
2. Multiply by W: We multiply by W (and add b) and pass through the ReLU
(or other) activation function to get the hidden layer h.
…
1 ^y p(aardvark|…)
and 0
0 1
1 35
…
thanks h1
0
wt-3 0
|V|
^y p(do|…)
for E 34
1 h2
…
0
0
y^42
0
1 992
all wt-2 p(fish|…)
0
0 h3
…
|V|
E
^y35102
…
the 1
wt-1 0
0
hdh
…
0
? 1 451
wt 0 ^y p(zebra|…)
0
|V|
E e W h U |V|
…
x d⨉|V| 3d⨉1 dh⨉3d dh⨉1 |V|⨉dh y
|V|⨉3 |V|⨉1
input layer embedding hidden output layer

one-hot layer layer softmax
vectors
Figure 7.13 Forward inference in a feedforward neural language model. At each timestep
t the network computes a d-dimensional embedding for each context word (by multiplying a
one-hot vector by the embedding matrix E), and concatenates the 3 resulting embeddings to
get the embedding layer e. The embedding vector e is multiplied by a weight matrix W and
then an activation function is applied element-wise to produce the hidden layer h, which is
then multiplied by another weight matrix U. Finally, a softmax output layer predicts at each
node i the probability that the next word wt will be vocabulary word Vi .
3. Multiply by U: h is now multiplied by U

4. Apply softmax: After the softmax, each node i in the output layer estimates
the probability P(wt = i|wt−1 , wt−2 , wt−3 )
In summary, the equations for a neural language model with a window size of 3,
given one-hot input vectors for each input context word, are:
e = [Ext−3 ; Ext−2 ; Ext−1 ]

h = σ (We + b)
z = Uh
Note that we formed the embedding layer e by concatenating the 3 embeddings

for the three context vectors; we’ll often use semicolons to mean concatenation of
vectors.
In the next section we’ll introduce a general algorithm for training neural net-
works, and then return to how to specifically train the neural language model in
Section 7.7.
7.6 • T RAINING N EURAL N ETS 17
7.6 Training Neural Nets

A feedforward neural net is an instance of supervised machine learning in which we
know the correct output y for each observation x. What the system produces, via
Eq. 7.13, is ŷ, the system’s estimate of the true y. The goal of the training procedure
is to learn parameters W[i] and b[i] for each layer i that make ŷ for each training
observation as close as possible to the true y.
In general, we do all this by drawing on the methods we introduced in Chapter 5
for logistic regression, so the reader should be comfortable with that chapter before
proceeding.
First, we’ll need a loss function that models the distance between the system
output and the gold output, and it’s common to use the loss function used for logistic
regression, the cross-entropy loss.
Second, to find the parameters that minimize this loss function, we’ll use the
gradient descent optimization algorithm introduced in Chapter 5.
Third, gradient descent requires knowing the gradient of the loss function, the
vector that contains the partial derivative of the loss function with respect to each
of the parameters. In logistic regression, for each observation we could directly
compute the derivative of the loss function with respect to an individual w or b. But
for neural networks, with millions of parameters in many layers, it’s much harder to
see how to compute the partial derivative of some weight in layer 1 when the loss
is attached to some much later layer. How do we partial out the loss over all those
intermediate layers? The answer is the algorithm called error backpropagation or
backward differentiation.
7.6.1 Loss function

cross-entropy The cross-entropy loss that is used in neural networks is the same one we saw for
loss
logistic regression. If the neural network is being used as a binary classifier, with
the sigmoid at the final layer, the loss function is the same logistic regression loss
we saw in Eq. ??:
LCE (ŷ, y) = − log p(y|x) = − [y log ŷ + (1 − y) log(1 − ŷ)] (7.25)
If we are using the network to classify into 3 or more classes, the loss function is
exactly the same as the loss for multinomial regression that we saw in Chapter 5 on
page ??. Let’s briefly summarize the explanation here for convenience. First, when
we have more than 2 classes we’ll need to represent both y and ŷ as vectors. Let’s
assume we’re doing hard classification, where only one class is the correct one.
The true label y is then a vector with K elements, each corresponding to a class,
with yc = 1 if the correct class is c, with all other elements of y being 0. Recall that
a vector like this, with one value equal to 1 and the rest 0, is called a one-hot vector.
And our classifier will produce an estimate vector with K elements ŷ, each element
ŷk of which represents the estimated probability p(yk = 1|x).
The loss function for a single example x is the negative sum of the logs of the K
output classes, each weighted by their probability yk :
K
X
LCE (ŷ, y) = − yk log ŷk (7.26)
k=1
We can simplify this equation further; let’s first rewrite the equation using the func-
tion 1{} which evaluates to 1 if the condition in the brackets is true and to 0 oth-
erwise. This makes it more obvious that the terms in the sum in Eq. 7.26 will be 0
except for the term corresponding to the true class for which yk = 1:
K
1{yk = 1} log ŷk
X
LCE (ŷ, y) = −
k=1
In other words, the cross-entropy loss is simply the negative log of the output proba-
bility corresponding to the correct class, and we therefore also call this the negative
negative log log likelihood loss:
likelihood loss
LCE (ŷ, y) = − log ŷc (where c is the correct class) (7.27)
Plugging in the softmax formula from Eq. 7.9, and with K the number of classes:
exp(zc )
LCE (ŷ, y) = − log PK (where c is the correct class) (7.28)
j=1 exp(z j )
7.6.2 Computing the Gradient

How do we compute the gradient of this loss function? Computing the gradient
requires the partial derivative of the loss function with respect to each parameter.
For a network with one weight layer and sigmoid output (which is what logistic
regression is), we could simply use the derivative of the loss that we used for logistic
regression in Eq. 7.29 (and derived in Section ??):
∂ LCE (ŷ, y)
= (ŷ − y) x j
∂wj
= (σ (w · x + b) − y) x j (7.29)
Or for a network with one weight layer and softmax output (=multinomial logistic
regression), we could use the derivative of the softmax loss from Eq. ??, shown for
a particular weight wk and input xi
∂ LCE (ŷ, y)
= −(yk − ŷk )xi
∂ wk,i
= −(yk − p(yk = 1|x))xi
!
exp (wk · x + bk )
= − yk − PK xi (7.30)
j=1 exp (wj · x + b j )
But these derivatives only give correct updates for one weight layer: the last one!
For deep networks, computing the gradients for each weight is much more complex,
since we are computing the derivative with respect to weight parameters that appear
all the way back in the very early layers of the network, even though the loss is
computed only at the very end of the network.
The solution to computing this gradient is an algorithm called error backprop-
error back-
propagation agation or backprop (Rumelhart et al., 1986). While backprop was invented spe-
cially for neural networks, it turns out to be the same as a more general procedure
called backward differentiation, which depends on the notion of computation
graphs. Let’s see how that works in the next subsection.
7.6.3 Computation Graphs

A computation graph is a representation of the process of computing a mathematical
expression, in which the computation is broken down into separate operations, each
of which is modeled as a node in a graph.
Consider computing the function L(a, b, c) = c(a + 2b). If we make each of the
component addition and multiplication operations explicit, and add names (d and e)
for the intermediate outputs, the resulting series of computations is:
d = 2∗b
e = a+d
L = c∗e
We can now represent this as a graph, with nodes for each operation, and di-
rected edges showing the outputs from each operation as the inputs to the next, as
in Fig. 7.14. The simplest use of computation graphs is to compute the value of
the function with some given inputs. In the figure, we’ve assumed the inputs a = 3,
b = 1, c = −2, and we’ve shown the result of the forward pass to compute the re-
sult L(3, 1, −2) = −10. In the forward pass of a computation graph, we apply each
operation left to right, passing the outputs of each computation as the input to the
next node.
forward pass
a a=3
e=a+d e=5
d=2
b=1
b d = 2b L=ce L=-10
c=-2
c
Figure 7.14 Computation graph for the function L(a, b, c) = c(a+2b), with values for input
nodes a = 3, b = 1, c = −2, showing the forward pass computation of L.
7.6.4 Backward differentiation on computation graphs

The importance of the computation graph comes from the backward pass, which
is used to compute the derivatives that we’ll need for the weight update. In this
example our goal is to compute the derivative of the output function L with respect
to each of the input variables, i.e., ∂∂ La , ∂∂ Lb , and ∂∂ Lc . The derivative ∂∂ La tells us how
much a small change in a affects L.
chain rule Backwards differentiation makes use of the chain rule in calculus, so let’s re-
mind ourselves of that. Suppose we are computing the derivative of a composite
function f (x) = u(v(x)). The derivative of f (x) is the derivative of u(x) with respect
to v(x) times the derivative of v(x) with respect to x:
df du dv
= · (7.31)
dx dv dx
The chain rule extends to more than two functions. If computing the derivative of a
composite function f (x) = u(v(w(x))), the derivative of f (x) is:
df du dv dw
= · · (7.32)
dx dv dw dx
The intuition of backward differentiation is to pass gradients back from the final
node to all the nodes in the graph. Fig. 7.15 shows part of the backward computation
at one node e. Each node takes an upstream gradient that is passed in from its parent
node to the right, and for each of its inputs computes a local gradient (the gradient
of its output with respect to its input), and uses the chain rule to multiply these two
to compute a downstream gradient to be passed on to the next earlier node.
d e
d e L
∂L = ∂L ∂e ∂e ∂L
∂d ∂e ∂d ∂d ∂e
downstream local upstream
gradient gradient gradient
Figure 7.15 Each node (like e here) takes an upstream gradient, multiplies it by the local
gradient (the gradient of its output with respect to its input), and uses the chain rule to compute
a downstream gradient to be passed on to a prior node. A node may have multiple local
gradients if it has multiple inputs.
Let’s now compute the 3 derivatives we need. Since in the computation graph
L = ce, we can directly compute the derivative ∂∂ Lc :
∂L
=e (7.33)
∂c
For the other two, we’ll need to use the chain rule:
∂L ∂L ∂e
=
∂a ∂e ∂a
∂L ∂L ∂e ∂d
= (7.34)
∂b ∂e ∂d ∂b
Eq. 7.34 and Eq. 7.33 thus require five intermediate derivatives: ∂∂ Le , ∂∂ Lc , ∂∂ ae , ∂∂ de , and
∂d
∂ b , which are as follows (making use of the fact that the derivative of a sum is the
sum of the derivatives):
∂L ∂L
L = ce : = c, =e
∂e ∂c
∂e ∂e
e = a+d : = 1, =1
∂a ∂d
∂d
d = 2b : =2
∂b
In the backward pass, we compute each of these partials along each edge of the
graph from right to left, using the chain rule just as we did above. Thus we begin by
computing the downstream gradients from node L, which are ∂∂ Le and ∂∂ Lc . For node e,
we then multiply this upstream gradient ∂∂ Le by the local gradient (the gradient of the
output with respect to the input), ∂∂ de to get the output we send back to node d: ∂∂ Ld .
And so on, until we have annotated the graph all the way to all the input variables.
The forward pass conveniently already will have computed the values of the forward
intermediate variables we need (like d and e) to compute these derivatives. Fig. 7.16
shows the backward pass.
a=3
a
∂L = ∂L ∂e =-2
∂a ∂e ∂a e=5
e=d+a
d=2
b=1 ∂e ∂e
=1 =1 ∂L
b d = 2b ∂L = ∂L ∂e =-2 ∂a ∂d =-2 L=-10
∂e L=ce
∂L = ∂L ∂d =-4 ∂d ∂d ∂e ∂d
=2 ∂L
∂b ∂d ∂b ∂b =-2
∂e
c=-2 ∂L
=5
∂c
∂L =5 backward pass
c ∂c
Figure 7.16 Computation graph for the function L(a, b, c) = c(a + 2b), showing the backward pass computa-
tion of ∂∂ La , ∂∂ Lb , and ∂∂ Lc .
Backward differentiation for a neural network

Of course computation graphs for real neural networks are much more complex.
Fig. 7.17 shows a sample computation graph for a 2-layer neural network with n0 =
2, n1 = 2, and n2 = 1, assuming binary classification and hence using a sigmoid
output unit for simplicity. The function that the computation graph is computing is:
z[1] = W[1] x + b[1]

a[1] = ReLU(z[1] )
z[2] = W[2] a[1] + b[2]
a[2] = σ (z[2] )
ŷ = a[2] (7.35)
For the backward pass we’ll also need to compute the loss L. The loss function
for binary sigmoid output from Eq. 7.25 is
LCE (ŷ, y) = − [y log ŷ + (1 − y) log(1 − ŷ)] (7.36)
Our output ŷ = a[2] , so we can rephrase this as

h i
LCE (a[2] , y) = − y log a[2] + (1 − y) log(1 − a[2] ) (7.37)
The weights that need updating (those for which we need to know the partial
derivative of the loss function) are shown in teal. In order to do the backward pass,
we’ll need to know the derivatives of all the functions in the graph. We already saw
in Section ?? the derivative of the sigmoid σ :
dσ (z)
= σ (z)(1 − σ (z)) (7.38)
dz
We’ll also need the derivatives of each of the other activation functions. The
derivative of tanh is:
d tanh(z)
= 1 − tanh2 (z) (7.39)
dz
[1]
w11
*
w[1]
12 z[1] = a1[1] =
* 1
+ ReLU
x1
*
b[1]
1 z[2] =
w[2] a[2] = σ L (a[2],y)
x2 11 +
*
*
w[1] z[1] a[1] w[2]

21 * 2 = 2 = 12
+ ReLU
w[1] b[2]
22 1
b[1]
2
Figure 7.17 Sample computation graph for a simple 2-layer neural net (= 1 hidden layer) with two input units
and 2 hidden units. We’ve adjusted the notation a bit to avoid long equations in the nodes by just mentioning
[1]
the function that is being computed, and the resulting variable name. Thus the * to the right of node w11 means
[1]
that w11 is to be multiplied by x1 , and the node z[1] = + means that the value of z[1] is computed by summing
[1]
the three nodes that feed into it (the two products, and the bias term bi ).
The derivative of the ReLU is

d ReLU(z) 0 f or z < 0
= (7.40)
dz 1 f or z ≥ 0
We’ll give the start of the computation, computing the derivative of the loss
function L with respect to z, or ∂∂Lz (and leaving the rest of the computation as an
exercise for the reader). By the chain rule:
∂L ∂ L ∂ a[2]
= [2] (7.41)
∂z ∂a ∂z
∂L
So let’s first compute ∂ a[2]
, taking the derivative of Eq. 7.37, repeated here:
h i
LCE (a[2] , y) = − y log a[2] + (1 − y) log(1 − a[2] )
! !
∂L ∂ log(a[2] ) ∂ log(1 − a[2] )
= − y + (1 − y)
∂ a[2] ∂ a[2] ∂ a[2]

1 1
= − y [2] + (1 − y) (−1)
a 1 − a[2]

y y−1
= − [2] + (7.42)
a 1 − a[2]
Next, by the derivative of the sigmoid:
∂ a[2]
= a[2] (1 − a[2] )
∂z
7.7 • T RAINING THE NEURAL LANGUAGE MODEL 23
Finally, we can use the chain rule:
∂L ∂ L ∂ a[2]
=
∂z ∂ a[2] ∂ z
y y−1
= − [2] + a[2] (1 − a[2] )
a 1 − a[2]
= a[2] − y (7.43)
Continuing the backward computation of the gradients (next by passing the gra-
[2]
dients over b1 and the two product nodes, and so on, back to all the teal nodes), is
left as an exercise for the reader.
7.6.5 More details on learning

Optimization in neural networks is a non-convex optimization problem, more com-
plex than for logistic regression, and for that and other reasons there are many best
practices for successful learning.
For logistic regression we can initialize gradient descent with all the weights and
biases having the value 0. In neural networks, by contrast, we need to initialize the
weights with small random numbers. It’s also helpful to normalize the input values
to have 0 mean and unit variance.
Various forms of regularization are used to prevent overfitting. One of the most
dropout important is dropout: randomly dropping some units and their connections from
the network during training (Hinton et al. 2012, Srivastava et al. 2014). Tuning
hyperparameter of hyperparameters is also important. The parameters of a neural network are the
weights W and biases b; those are learned by gradient descent. The hyperparameters
are things that are chosen by the algorithm designer; optimal values are tuned on a
devset rather than by gradient descent learning on the training set. Hyperparameters
include the learning rate η, the mini-batch size, the model architecture (the number
of layers, the number of hidden nodes per layer, the choice of activation functions),
how to regularize, and so on. Gradient descent itself also has many architectural
variants such as Adam (Kingma and Ba, 2015).
Finally, most modern neural networks are built using computation graph for-
malisms that make it easy and natural to do gradient computation and parallelization
on vector-based GPUs (Graphic Processing Units). PyTorch (Paszke et al., 2017)
and TensorFlow (Abadi et al., 2015) are two of the most popular. The interested
reader should consult a neural network textbook for further details; some sugges-
tions are at the end of the chapter.
7.7 Training the neural language model

Now that we’ve seen how to train a generic neural net, let’s talk about the architec-
ture for training a neural language model, setting the parameters θ = E, W, U, b.
freeze For some tasks, it’s ok to freeze the embedding layer E with initial word2vec val-
ues. Freezing means we use word2vec or some other pretraining algorithm to com-
pute the initial embedding matrix E, and then hold it constant while we only modify
W, U, and b, i.e., we don’t update E during language model training. However, often
we’d like to learn the embeddings simultaneously with training the network. This is
useful when the task the network is designed for (like sentiment classification, trans-
lation, or parsing) places strong constraints on what makes a good representation for
words.
Let’s see how to train the entire model including E, i.e. to set all the parameters
θ = E, W, U, b. We’ll do this via gradient descent (Fig. ??), using error backpropa-
gation on the computation graph to compute the gradient. Training thus not only sets
the weights W and U of the network, but also as we’re predicting upcoming words,
we’re learning the embeddings E for each word that best predict upcoming words.
wt=fish L = −log P(fish | for, all, the)

…
1 ^y p(aardvark|…)
and 0
0 1
1 35
…
thanks h1
0
wt-3 0
|V|
^y p(do|…)
for E 34
1 h2
…
0
0
y^42
0
1 992
all wt-2 p(fish|…)
0
0 h3
…
|V|
E
^y35102
…
the 1
wt-1 0
0
hdh
…
0
fish 1 451
wt 0 ^y p(zebra|…)
0
|V|
E e W h U |V|
…
x d⨉|V| 3d⨉1 dh⨉3d dh⨉1 |V|⨉dh y
|V|⨉3 |V|⨉1
input layer embedding hidden output layer

one-hot layer layer softmax
vectors
Figure 7.18 Learning all the way back to embeddings. Again, the embedding matrix E is
shared among the 3 context words.
Fig. 7.18 shows the set up for a window size of N=3 context words. The input x
consists of 3 one-hot vectors, fully connected to the embedding layer via 3 instanti-
ations of the embedding matrix E. We don’t want to learn separate weight matrices
for mapping each of the 3 previous words to the projection layer. We want one single
embedding dictionary E that’s shared among these three. That’s because over time,
many different words will appear as wt−2 or wt−1 , and we’d like to just represent
each word with one vector, whichever context position it appears in. Recall that the
embedding weight matrix E has a column for each word, each a column vector of d
dimensions, and hence has dimensionality d × |V |.
Generally training proceeds by taking as input a very long text, concatenating all
the sentences, starting with random weights, and then iteratively moving through the
text predicting each word wt . At each word wt , we use the cross-entropy (negative
log likelihood) loss. Recall that the general form for this (repeated from Eq. 7.27 is:
LCE (ŷ, y) = − log ŷi , (where i is the correct class) (7.44)
For language modeling, the classes are the words in the vocabulary, so ŷi here means
the probability that the model assigns to the correct next word wt :
LCE = − log p(wt |wt−1 , ..., wt−n+1 ) (7.45)
7.8 • S UMMARY 25
The parameter update for stochastic gradient descent for this loss from step s to s + 1
is then:
∂ [− log p(wt |wt−1 , ..., wt−n+1 )]
θ s+1 = θ s − η (7.46)
∂θ
This gradient can be computed in any standard neural network framework which
will then backpropagate through θ = E, W, U, b.
Training the parameters to minimize loss will result both in an algorithm for
language modeling (a word predictor) but also a new set of embeddings E that can
be used as word representations for other tasks.
7.8 Summary
• Neural networks are built out of neural units, originally inspired by human
neurons but now simply an abstract computational device.
• Each neural unit multiplies input values by a weight vector, adds a bias, and
then applies a non-linear activation function like sigmoid, tanh, or rectified
linear unit.
• In a fully-connected, feedforward network, each unit in layer i is connected
to each unit in layer i + 1, and there are no cycles.
• The power of neural networks comes from the ability of early layers to learn
representations that can be utilized by later layers in the network.
• Neural networks are trained by optimization algorithms like gradient de-
scent.
• Error backpropagation, backward differentiation on a computation graph,
is used to compute the gradients of the loss function for a network.
• Neural language models use a neural network as a probabilistic classifier, to
compute the probability of the next word given the previous n words.
• Neural language models can use pretrained embeddings, or can learn embed-
dings from scratch in the process of language modeling.
Bibliographical and Historical Notes

The origins of neural networks lie in the 1940s McCulloch-Pitts neuron (McCul-
loch and Pitts, 1943), a simplified model of the human neuron as a kind of com-
puting element that could be described in terms of propositional logic. By the late
1950s and early 1960s, a number of labs (including Frank Rosenblatt at Cornell and
Bernard Widrow at Stanford) developed research into neural networks; this phase
saw the development of the perceptron (Rosenblatt, 1958), and the transformation
of the threshold into a bias, a notation we still use (Widrow and Hoff, 1960).
The field of neural networks declined after it was shown that a single perceptron
unit was unable to model functions as simple as XOR (Minsky and Papert, 1969).
While some small amount of work continued during the next two decades, a major
revival for the field didn’t come until the 1980s, when practical tools for building
deeper networks like error backpropagation became widespread (Rumelhart et al.,
1986). During the 1980s a wide variety of neural network and related architec-
tures were developed, particularly for applications in psychology and cognitive sci-
ence (Rumelhart and McClelland 1986b, McClelland and Elman 1986, Rumelhart
connectionist and McClelland 1986a, Elman 1990), for which the term connectionist or paral-
lel distributed processing was often used (Feldman and Ballard 1982, Smolensky
1988). Many of the principles and techniques developed in this period are foun-
dational to modern work, including the ideas of distributed representations (Hinton,
1986), recurrent networks (Elman, 1990), and the use of tensors for compositionality
(Smolensky, 1990).
By the 1990s larger neural networks began to be applied to many practical lan-
guage processing tasks as well, like handwriting recognition (LeCun et al. 1989) and
speech recognition (Morgan and Bourlard 1990). By the early 2000s, improvements
in computer hardware and advances in optimization and training techniques made it
possible to train even larger and deeper networks, leading to the modern term deep
learning (Hinton et al. 2006, Bengio et al. 2007). We cover more related history in
Chapter 9 and Chapter 16.
There are a number of excellent books on the subject. Goldberg (2017) has
superb coverage of neural networks for natural language processing. For neural
networks in general see Goodfellow et al. (2016) and Nielsen (2015).
Bibliographical and Historical Notes 27
Abadi, M., A. Agarwal, P. Barham, E. Brevdo, Z. Chen, Rumelhart, D. E., G. E. Hinton, and R. J. Williams. 1986.
C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, Learning internal representations by error propagation. In
S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Is- D. E. Rumelhart and J. L. McClelland, editors, Parallel
ard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Leven- Distributed Processing, volume 2, pages 318–362. MIT
berg, D. Mané, R. Monga, S. Moore, D. Murray, C. Olah, Press.
M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Tal- Rumelhart, D. E. and J. L. McClelland. 1986a. On learning
war, P. Tucker, V. Vanhoucke, V. Vasudevan, F. Viégas, the past tense of English verbs. In D. E. Rumelhart and
O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, J. L. McClelland, editors, Parallel Distributed Process-
and X. Zheng. 2015. TensorFlow: Large-scale machine ing, volume 2, pages 216–271. MIT Press.
learning on heterogeneous systems. Software available
from tensorflow.org. Rumelhart, D. E. and J. L. McClelland, editors. 1986b. Par-
allel Distributed Processing. MIT Press.
Bengio, Y., R. Ducharme, P. Vincent, and C. Jauvin. 2003.
A neural probabilistic language model. JMLR, 3:1137– Russell, S. and P. Norvig. 2002. Artificial Intelligence: A
1155. Modern Approach, 2nd edition. Prentice Hall.
Bengio, Y., P. Lamblin, D. Popovici, and H. Larochelle. Smolensky, P. 1988. On the proper treatment of connection-
2007. Greedy layer-wise training of deep networks. ism. Behavioral and brain sciences, 11(1):1–23.
NeurIPS. Smolensky, P. 1990. Tensor product variable binding and
Elman, J. L. 1990. Finding structure in time. Cognitive sci- the representation of symbolic structures in connectionist
ence, 14(2):179–211. systems. Artificial intelligence, 46(1-2):159–216.
Feldman, J. A. and D. H. Ballard. 1982. Connectionist mod- Srivastava, N., G. E. Hinton, A. Krizhevsky, I. Sutskever,
els and their properties. Cognitive Science, 6:205–254. and R. R. Salakhutdinov. 2014. Dropout: a simple
way to prevent neural networks from overfitting. JMLR,
Goldberg, Y. 2017. Neural Network Methods for Natural 15(1):1929–1958.
Language Processing, volume 10 of Synthesis Lectures
on Human Language Technologies. Morgan & Claypool. Widrow, B. and M. E. Hoff. 1960. Adaptive switching cir-
cuits. IRE WESCON Convention Record, volume 4.
Goodfellow, I., Y. Bengio, and A. Courville. 2016. Deep
Learning. MIT Press.
Hinton, G. E. 1986. Learning distributed representations of
concepts. COGSCI.
Hinton, G. E., S. Osindero, and Y.-W. Teh. 2006. A fast
learning algorithm for deep belief nets. Neural computa-
tion, 18(7):1527–1554.
Hinton, G. E., N. Srivastava, A. Krizhevsky, I. Sutskever, and
R. R. Salakhutdinov. 2012. Improving neural networks
by preventing co-adaptation of feature detectors. ArXiv
preprint arXiv:1207.0580.
Kingma, D. and J. Ba. 2015. Adam: A method for stochastic
optimization. ICLR 2015.
LeCun, Y., B. Boser, J. S. Denker, D. Henderson, R. E.
Howard, W. Hubbard, and L. D. Jackel. 1989. Backprop-
agation applied to handwritten zip code recognition. Neu-
ral computation, 1(4):541–551.
McClelland, J. L. and J. L. Elman. 1986. The TRACE model
of speech perception. Cognitive Psychology, 18:1–86.
McCulloch, W. S. and W. Pitts. 1943. A logical calculus of
ideas immanent in nervous activity. Bulletin of Mathe-
matical Biophysics, 5:115–133.
Minsky, M. and S. Papert. 1969. Perceptrons. MIT Press.
Morgan, N. and H. Bourlard. 1990. Continuous speech
recognition using multilayer perceptrons with hidden
markov models. ICASSP.
Nielsen, M. A. 2015. Neural networks and Deep learning.
Determination Press USA.
Paszke, A., S. Gross, S. Chintala, G. Chanan, E. Yang, Z. De-
Vito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer.
2017. Automatic differentiation in pytorch. NIPS-W.
Rosenblatt, F. 1958. The perceptron: A probabilistic model
for information storage and organization in the brain. Psy-
chological review, 65(6):386–408.

Neural Networks and Neural Language Models

Uploaded by

Copyright:

Available Formats

Neural Networks and Neural Language Models

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Neural Networks and Neural Language Models

Uploaded by

Copyright:

Available Formats

Speech and Language Processing. Daniel Jurafsky & James H. Martin. Copyright © 2023.

Neural Networks and Neural

“[M]achines of this character can behave in a very complicated manner when

Neural networks are a fundamental computational tool for language process-

w = [0.2, 0.3, 0.9]

What would this unit do with the following input vector:

x = [0.5, 0.6, 0.1]

The resulting output y would be:

y = ReLU(z) = max(z, 0) (7.6)

7.2 The XOR problem

7.2.1 The solution: neural networks

a) The original x space b) The new (linearly separable) h space

7.3 Feedforward Neural Networks

Thus for example given a vector

the softmax function will normalize it to a probability distribution (shown rounded):

7.3.1 More details on feedforward networks

z[1] = W[1] a[0] + b[1]

z[1] = W[1] x + b[1]

We can rewrite the function that the network is computing as:

z[2] = W[2] z[1] + b[2]

This generalizes to any number of layers. So without non-linear activation functions,

we’ll instead use:

7.4 Feedforward networks for NLP: Classification

great count of “no” x3 p(neut)

Most applications of neural networks for NLP do something different, however.

x = mean(e(w1 ), e(w2 ), . . . , e(wn ))

embedding for ^y p(neut)

Input layer Hidden layer Output layer

The idea of using word2vec or GloVe embeddings as our input representation—

7.5 Feedforward Neural Language Modeling

P(wt |w1 , . . . , wt−1 ) ≈ P(wt |wt−N+1 , . . . , wt−1 ) (7.23)

7.5.1 Forward inference in the neural language model

The 3 resulting embedding vectors are concatenated to produce e, the embedding

input layer embedding hidden output layer

3. Multiply by U: h is now multiplied by U

e = [Ext−3 ; Ext−2 ; Ext−1 ]

Note that we formed the embedding layer e by concatenating the 3 embeddings

7.6 Training Neural Nets

7.6.1 Loss function

LCE (ŷ, y) = − log ŷc (where c is the correct class) (7.27)

7.6.2 Computing the Gradient

7.6.3 Computation Graphs

7.6.4 Backward differentiation on computation graphs

Backward differentiation for a neural network

z[1] = W[1] x + b[1]

LCE (ŷ, y) = − [y log ŷ + (1 − y) log(1 − ŷ)] (7.36)

Our output ŷ = a[2] , so we can rephrase this as

w[1] z[1] a[1] w[2]

The derivative of the ReLU is

Next, by the derivative of the sigmoid:

Finally, we can use the chain rule:

7.6.5 More details on learning

7.7 Training the neural language model

wt=fish L = −log P(fish | for, all, the)

input layer embedding hidden output layer

Bibliographical and Historical Notes

You might also like