June 2016 - This notebook was created by [Jordi Vitrià](http://www.ub.edu/cvub/jordivitria/). Source and [license](./LICENSE.txt) info are in the folder.
In [3]:
import warnings
warnings.filterwarnings('ignore')
import numpy as np
import seaborn as sns
import bokeh.plotting as bp
import matplotlib.pyplot as plt
from sklearn.datasets.samples_generator import make_regression
from scipy import stats
from bokeh.models import WheelZoomTool, ResetTool, PanTool
%matplotlib inline
In general Learning from Data is a scientific discipline that is concerned with the design and development of algorithms that allow computers to infer (from data) a model that allows compact representation (unsupervised learning) and/or good generalization (supervised learning).
This is an important technology because it enables computational systems to adaptively improve their performance with experience accumulated from the observed data.
Most of these algorithms are based on the iterative solution of a mathematical problem that involves data and model. If there was an analytical solution to the problem, this should be the adopted one, but this is not the case for most of the cases.
So, the most common strategy for learning from data is based on solving a system of equations as a way to find a series of parameters of the model that minimizes a mathematical problem. This is called optimization.
The most important technique for solving optimization problems is gradient descend.
See "An Interactive Tutorial on Numerical Optimization": http://www.benfrederickson.com/numerical-optimization/
The most simple thing we can try to minimize a function $f(x)$ would be to sample two points relatively near each other, and just repeatedly take a step down away from the largest value.
The Nelder-Mead method dynamically adjusts the step size based off the loss of the new point. If the new point is better than any previously seen value, it expands the step size to accelerate towards the bottom. Likewise if the new point is worse it contracts the step size to converge around the minima. The usual settings are to half the step size when contracting and double the step size when expanding.
This method can be easily extended into higher dimensional examples, all thats required is taking one more point than there are dimensions - and then reflecting the worst point around the rest of the points to take a step down.
Let's suppose that we have a function $f: \Re \rightarrow \Re$. For example:
$$f(x) = x^2$$Our objective is to find the argument $x$ that minimizes this function (for maximization, consider $-f(x)$). To this end, the critical concept is the derivative.
The derivative of $f$ of a variable $x$, $f'(x)$ or $\frac{\mathrm{d}f}{\mathrm{d}x}$, is a measure of the rate at which the value of the function changes with respect to the change of the variable. It is defined as the following limit:
$$ f'(x) = \lim_{h \rightarrow 0} \frac{f(x + h) - f(x)}{h} $$The derivative specifies how to scale a small change in the input in order to obtain the corresponding change in the output:
$$ f(x + h) \approx f(x) + h f'(x)$$
In [4]:
# numerical derivative at a point x
def f(x):
return x**2
def fin_dif(x, f, h = 0.00001):
'''
This method returns the derivative of f at x
by using the finite difference method
'''
return (f(x+h) - f(x))/h
x = 2.0
print "{:2.4f}".format(fin_dif(x,f))
The limit as $h$ approaches zero, if it exists, should represent the slope of the tangent line to $(x, f(x))$.
For values that are not zero it is only an approximation.
In [5]:
for h in np.linspace(0.0, 1.0 , 5):
print "{:3.6f}".format(f(5+h)), "{:3.6f}".format(f(5)+h*fin_dif(5,f))
In [6]:
x = np.linspace(-1.5,-0.5, 100)
f = [i**2 for i in x]
plt.plot(x,f, 'r-')
plt.plot([-1.5, -0.5], [2, 0.0], 'k-', lw=2)
plt.plot([-1.4, -1.0], [1.96, 1.0], 'b-', lw=2)
plt.plot([-1],[1],'o')
plt.plot([-1.4],[1.96],'o')
plt.text(-1.0, 1.2, r'$x,f(x)$')
plt.text(-1.4, 2.2, r'$(x-h),f(x-h)$')
plt.gcf().set_size_inches((12,6))
plt.grid()
plt.show
Out[6]:
It can be shown that the “centered difference formula" is better when computing numerical derivatives:
$$ \lim_{h \rightarrow 0} \frac{f(x + h) - f(x - h)}{2h} $$The error in the "finite difference" approximation can be derived from Taylor's theorem and, assuming that $f$ is differentiable, is $O(h)$. In the case of “centered difference" the error is $O(h^2)$.
The derivative tells how to chage $x$ in order to make a small improvement in $f$.
Then, we can follow these steps to decrease the value of the function:
The search for the minima ends when the derivative is zero because we have no more information about which direction to move. $x$ is a critical o stationary point if $f'(x)=0$.
If $f$ is a convex function, this should be the minimum (maximum) of our functions. In other cases it could be a local minimum (maximum) or a saddle point.
In [10]:
W = 400
H = 250
bp.output_notebook()
In [11]:
x = np.linspace(-15,15,100)
y = x**2
TOOLS = [WheelZoomTool(), ResetTool(), PanTool()]
s1 = bp.figure(width=W, plot_height=H,
title='Local minimum of function',
tools=TOOLS)
s1.line(x, y, color="navy", alpha=0.5, line_width=3)
s1.circle(0, 0, size =10, color="orange")
s1.title_text_font_size = '12pt'
s1.yaxis.axis_label_text_font_size = "14pt"
s1.xaxis.axis_label_text_font_size = "14pt"
bp.show(s1)
In [12]:
x = np.linspace(-15,15,100)
y = -x**2
TOOLS = [WheelZoomTool(), ResetTool(), PanTool()]
s1 = bp.figure(width=W, plot_height=H,
title='Local maximum of function',
tools=TOOLS)
s1.line(x, y, color="navy", alpha=0.5, line_width=3)
s1.circle(0, 0, size =10, color="orange")
s1.title_text_font_size = '12pt'
s1.yaxis.axis_label_text_font_size = "14pt"
s1.xaxis.axis_label_text_font_size = "14pt"
bp.show(s1)
In [13]:
x = np.linspace(-15,15,100)
y = x**3
TOOLS = [WheelZoomTool(), ResetTool(), PanTool()]
s1 = bp.figure(width=W, plot_height=H,
title='Saddle point of function',
tools=TOOLS)
s1.line(x, y, color="navy", alpha=0.5, line_width=3)
s1.circle(0, 0, size =10, color="orange")
s1.title_text_font_size = '12pt'
s1.yaxis.axis_label_text_font_size = "14pt"
s1.xaxis.axis_label_text_font_size = "14pt"
bp.show(s1)
There are two problems with numerical derivatives:
Our knowledge from Calculus could help!
We know that we can get an analytical expression of the derivative for some functions.
For example, let's suppose we have a simple quadratic function, $f(x)=x^2−6x+5$, and we want to find the minimum of this function.
We can solve this analytically using Calculus, by finding the derivate $f'(x) = 2x-6$ and setting it to zero:
\begin{equation} \begin{split} 2x-6 & = & 0 \\ 2x & = & 6 \\ x & = & 3 \\ \end{split} \end{equation}
In [14]:
x = np.linspace(-10,20,100)
y = x**2 - 6*x + 5
TOOLS = [WheelZoomTool(), ResetTool(), PanTool()]
s1 = bp.figure(width=W, plot_height=H,
tools=TOOLS)
s1.line(x, y, color="navy", alpha=0.5, line_width=3)
s1.circle(3, 3**2 - 6*3 + 5, size =10, color="orange")
s1.title_text_font_size = '12pt'
s1.yaxis.axis_label_text_font_size = "14pt"
s1.xaxis.axis_label_text_font_size = "14pt"
bp.show(s1)
To find the local minimum using gradient descend: you start at a random point, and move into the direction of steepest descent relative to the derivative:
In this example, let's suppose we start at $x=15$. The derivative at this point is $2×15−6=24$.
Because we're using gradient descent, we need to subtract the gradient from our $x$-coordinate: $f(x - f'(x))$. However, notice that $15−24$ gives us $−9$, clearly overshooting over target of $3$.
In [15]:
x = np.linspace(-10,20,100)
y = x**2 - 6*x + 5
start = 15
TOOLS = [WheelZoomTool(), ResetTool(), PanTool()]
s1 = bp.figure(width=W, plot_height=H,
tools=TOOLS)
s1.line(x, y, color="navy", alpha=0.5, line_width=3)
s1.circle(start, start**2 - 6*start + 5, size =10, color="orange")
d = 2 * start - 6
end = start - d
s1.circle(end, end**2 - 6*end + 5, size =10, color="red")
s1.title_text_font_size = '12pt'
s1.yaxis.axis_label_text_font_size = "14pt"
s1.xaxis.axis_label_text_font_size = "14pt"
bp.show(s1)
To fix this, we multiply the gradient by a step size. This step size (often called alpha) has to be chosen carefully, as a value too small will result in a long computation time, while a value too large will not give you the right result (by overshooting) or even fail to converge.
In this example, we'll set the step size to 0.01, which means we'll subtract $24×0.01$ from $15$, which is $14.76$.
This is now our new temporary local minimum: We continue this method until we either don't see a change after we subtracted the derivative step size, or until we've completed a pre-set number of iterations.
In [16]:
old_min = 0
temp_min = 15
step_size = 0.01
precision = 0.0001
def f_derivative(x):
import math
return 2*x -6
mins = []
cost = []
while abs(temp_min - old_min) > precision:
old_min = temp_min
gradient = f_derivative(old_min)
move = gradient * step_size
temp_min = old_min - move
cost.append((3-temp_min)**2)
mins.append(temp_min)
# rounding the result to 2 digits because of the step size
print "Local minimum occurs at {:3.2f}.".format(round(temp_min,2))
An important feature of gradient descent is that there should be a visible improvement over time: In this example, we simply plotted the squared distance from the local minima calculated by gradient descent and the true local minimum, cost, against the iteration during which it was calculated. As we can see, the distance gets smaller over time, but barely changes in later iterations.
In [17]:
TOOLS = [WheelZoomTool(), ResetTool(), PanTool()]
x, y = (zip(*enumerate(cost)))
s1 = bp.figure(width=W,
height=H,
title='Squared distance to true local minimum',
# title_text_font_size='14pt',
tools=TOOLS,
x_axis_label = 'Iteration',
y_axis_label = 'Distance'
)
s1.line(x, y, color="navy", alpha=0.5, line_width=3)
s1.title_text_font_size = '16pt'
s1.yaxis.axis_label_text_font_size = "14pt"
s1.xaxis.axis_label_text_font_size = "14pt"
bp.show(s1)
Let's consider a $n$-dimensional function $f: \Re^n \rightarrow \Re$. For example:
$$f(\mathbf{x}) = \sum_{n} x_n^2$$Our objective is to find the argument $\mathbf{x}$ that minimizes this function.
The gradient of $f$ is the vector whose components are the $n$ partial derivatives of $f$. It is thus a vector-valued function.
The gradient points in the direction of the greatest rate of increase of the function.
$$\nabla {f} = (\frac{\partial f}{\partial x_1}, \dots, \frac{\partial f}{\partial x_n})$$
In [16]:
def f(x):
return sum(x_i**2 for x_i in x)
def fin_dif_partial_centered(x, f, i, h=1e-6):
w1 = [x_j + (h if j==i else 0) for j, x_j in enumerate(x)]
w2 = [x_j - (h if j==i else 0) for j, x_j in enumerate(x)]
return (f(w1) - f(w2))/(2*h)
def fin_dif_partial_old(x, f, i, h=1e-6):
w1 = [x_j + (h if j==i else 0) for j, x_j in enumerate(x)]
return (f(w1) - f(x))/h
def gradient_centered(x, f, h=1e-6):
return[round(fin_dif_partial_centered(x,f,i,h), 10) for i,_ in enumerate(x)]
def gradient_old(x, f, h=1e-6):
return[round(fin_dif_partial_old(x,f,i,h), 10) for i,_ in enumerate(x)]
x = [1.0,1.0,1.0]
print f(x), gradient_centered(x,f)
print f(x), gradient_old(x,f)
The function we have evaluated, $f({\mathbf x}) = x_1^2+x_2^2+x_3^2$, is $3$ at $(1,1,1)$ and the gradient vector at this point is $(2,2,2)$.
Then, we can follow this steps to maximize (or minimize) the function:
In [18]:
def euc_dist(v1,v2):
import numpy as np
import math
v = np.array(v1)-np.array(v2)
return math.sqrt(sum(v_i ** 2 for v_i in v))
Let's start by choosing a random vector and then walking a step in the opposite direction of the gradient vector. We will stop when the difference between the new solution and the old solution is less than a tolerance value.
In [19]:
# choosing a random vector
import random
import numpy as np
x = [random.randint(-10,10) for i in range(3)]
x
Out[19]:
In [19]:
def step(x,grad,alpha):
return [x_i - alpha * grad_i for x_i, grad_i in zip(x,grad)]
tol = 1e-15
alpha = 0.01
while True:
grad = gradient_centered(x,f)
next_x = step(x,grad,alpha)
if euc_dist(next_x,x) < tol:
break
x = next_x
print [round(i,10) for i in x]
The step size, alpha, is a slippy concept: if it is too small we will slowly converge to the solution, if it is too large we can diverge from the solution.
There are several policies to follow when selecting the step size:
The last policy is good, but too expensive. In this case we would consider a fixed set of values:
In [17]:
step_size = [100, 10, 1, 0.1, 0.01, 0.001, 0.0001, 0.00001]
In general, we have:
In the most common case $f$ represents the errors from a data representation model $M$. To fit the model is to find the optimal parameters $\mathbf{w}$ that minimize the following expression:
$$ f_\mathbf{w} = \sum_{i} (y_i - M(\mathbf{x}_i,\mathbf{w}))^2 $$For example, $(\mathbf{x},y)$ can represent:
If $y$ is a real value, it is called a regression problem.
If $y$ is binary/categorical, it is called a classification problem.
Let's suppose that $M(\mathbf{x},\mathbf{w}) = \mathbf{w} \cdot \mathbf{x}$.
We can implement gradient descend in the following way (batch gradient descend):
In [21]:
# f = 2x
x = range(100)
y = [2*i for i in x]
# f_target = Sum (y - wx)**2
def target_f(x,y,w):
import numpy as np
return np.sum((np.array(y) - np.array(x) * w)**2.0)
# gradient_f = Sum 2wx**2 - 2xy
def gradient_f(x,y,w):
import numpy as np
return np.sum(2*w*(np.array(x)**2) - 2*np.array(x)*np.array(y))
def step(w,grad,alpha):
return w - alpha * grad
def min_batch(target_f, gradient_f, x, y, toler = 1e-6):
import random
alphas = [100, 10, 1, 0.1, 0.001, 0.00001]
w = random.random()
val = target_f(x,y,w)
print "First w:", w, "First Val:", val, "\n"
i = 0
while True:
i += 1
gradient = gradient_f(x,y,w)
next_ws = [step(w, gradient, alpha) for alpha in alphas]
next_vals = [target_f(x,y,w) for w in next_ws]
min_val = min(next_vals)
next_w = next_ws[next_vals.index(min_val)]
next_val = target_f(x,y,next_w)
print i, "w: {:4.4f}".format(w), "Val:{:4.4f}".format(val), "Gradient:", gradient
if (abs(val - next_val) < toler) or (i>200):
return w
else:
w, val = next_w, next_val
min_batch(target_f, gradient_f, x, y)
Out[21]:
In [19]:
# Exercise:
# 1. Consider a set of 100 data points and explain the behavior of the algorithm.
# 2. How could we fix this behavior?
The last function evals the whole dataset $(\mathbf{x}_i,y_i)$ at every step.
If the dataset is large, this strategy is too costly. In this case we will use a strategy called SGD (Stochastic Gradient Descend).
When learning from data, the cost function is additive: it is computed by adding sample reconstruction errors.
Then, we can compute the estimate the gradient (and move towards the minimum) by using only one data sample (or a small data sample).
Thus, we will find the minimum by iterating this gradient estimation over the dataset.
A full iteration over the dataset is called epoch. During an epoch, data must be used in a random order.
If we apply this method we have some theoretical guarantees to find the minimum.
In [25]:
import numpy as np
x = range(10)
y = [2*i for i in x]
data = zip(x,y)
def in_random_order(data):
import random
indexes = [i for i,_ in enumerate(data)]
random.shuffle(indexes)
for i in indexes:
yield data[i]
for (x_i,y_i) in in_random_order(data):
print x_i,y_i
In [26]:
def gradient_f_SGD(x,y,w):
import numpy as np
return 2*w*(np.array(x)**2) - 2*np.array(x)*np.array(y)
def SGD(target_f, gradient_f, x, y, alpha_0=0.01):
import numpy as np
import random
data = zip(x,y)
w = random.random()
alpha = alpha_0
min_w, min_val = float('inf'), float('inf')
iteration_no_increase = 0
while iteration_no_increase < 100:
val = sum(target_f(x_i, y_i, w) for x_i,y_i in data)
if val < min_val:
min_w, min_val = w, val
iteration_no_increase = 0
alpha = alpha_0
else:
iteration_no_increase += 1
alpha *= 0.9
for x_i, y_i in in_random_order(data):
gradient_i = gradient_f(x_i, y_i, w)
w = np.array(w) - (alpha * np.array(gradient_i))
return min_w
In [27]:
print "w:", SGD(target_f, gradient_f_SGD, x, y, 0.01)
In [28]:
import numpy as np
x = np.random.uniform(0,1,20)
def f(x): return x*2
noise_variance =0.2
noise = np.random.randn(x.shape[0])*noise_variance
y = f(x) + noise
plt.plot(x, y, 'o', label='y')
plt.plot([0, 1], [f(0), f(1)], 'b-', label='f(x)')
plt.xlabel('$x$', fontsize=15)
plt.ylabel('$t$', fontsize=15)
plt.ylim([0,2])
plt.title('inputs (x) vs targets (y)')
plt.grid()
plt.legend(loc=2)
plt.gcf().set_size_inches((10,6))
plt.show()
In [29]:
# Our model y = x * w
def nn(x, w): return x * w
# Our cost function
def cost(y, t): return ((t - y)**2).sum()
ws = np.linspace(0, 4, num=100)
cost_ws = np.vectorize(lambda w: cost(nn(x, w) , y))(ws)
# Ploting the cost function
plt.plot(ws, cost_ws, 'r-')
plt.xlabel('$w$', fontsize=15)
plt.ylabel('Cost', fontsize=15)
plt.title('Cost vs. $w$')
plt.grid()
plt.gcf().set_size_inches((10,6))
plt.show()
Complete the following code and look at the plot of the first gradient descent updates. Explore the behavior of the proposed learning rates.
In [30]:
def gradient(w, x, y):
return 2 * x * (nn(x, w) - y)
def step(w_k, x, y, learning_rate):
return learning_rate * gradient(w_k, x, y).sum()
w = 0.01
# define a learning_rate
learning_rate = 0.1
nb_of_iterations = 20
w_cost = [(w, cost(nn(x, w), y))]
for i in range(nb_of_iterations):
# Here your code
w_cost.append((w, cost(nn(x, w), y)))
for i in range(0, len(w_cost)):
print('w({}): {:.4f} \t cost: {:.4f}'.format(i, w_cost[i][0], w_cost[i][1]))
In [31]:
# Plotting the first gradient descent updates
plt.plot(ws, cost_ws, 'r-') # Plot the error curve
# Plot the updates
for i in range(1, len(w_cost)-2):
w1, c1 = w_cost[i-1]
w2, c2 = w_cost[i]
plt.plot(w1, c1, 'bo')
plt.plot([w1, w2],[c1, c2], 'b-')
plt.text(w1, c1+0.5, '$w({})$'.format(i))
# Plot the last weight, axis, and show figure
w1, c1 = w_cost[len(w_cost)-3]
plt.plot(w1, c1, 'bo')
plt.text(w1, c1+0.5, '$w({})$'.format(nb_of_iterations))
plt.xlabel('$w$', fontsize=15)
plt.ylabel('$\\xi$', fontsize=15)
plt.title('Gradient descent updates plotted on cost function')
plt.grid()
plt.gcf().set_size_inches((10,6))
plt.show()
In [32]:
w = 0
nb_of_iterations = 10
for i in range(nb_of_iterations):
dw = step(w, x, y, learning_rate)
w = w - dw
plt.plot(x, y, 'o', label='t')
plt.plot([0, 1], [f(0), f(1)], 'b-', label='f(x)')
plt.plot([0, 1], [0*w, 1*w], 'r-', label='fitted line')
plt.xlabel('input x')
plt.ylabel('target t')
plt.ylim([0,2])
plt.title('input vs. target')
plt.grid()
plt.legend(loc=2)
plt.gcf().set_size_inches((10,6))
plt.show()
In code, general batch gradient descent looks something like this:
nb_epochs = 100
for i in range(nb_epochs):
grad = evaluate_gradient(target_f, data, w)
w = w - learning_rate * grad
For a pre-defined number of epochs, we first compute the gradient vector of the target function for the whole dataset w.r.t. our parameter vector.
Stochastic gradient descent (SGD) in contrast performs a parameter update for each training example and label:
nb_epochs = 100
for i in range(nb_epochs):
np.random.shuffle(data)
for example in data:
grad = evaluate_gradient(target_f, example, w)
w = w - learning_rate * grad
Mini-batch gradient descent finally takes the best of both worlds and performs an update for every mini-batch of $n$ training examples:
nb_epochs = 100
for i in range(nb_epochs):
np.random.shuffle(data)
for batch in get_batches(data, batch_size=50):
grad = evaluate_gradient(target_f, batch, w)
w = w - learning_rate * grad
Minibatch SGD has the advantage that it works with a slightly less noisy estimate of the gradient. However, as the minibatch size increases, the number of updates done per computation done decreases (eventually it becomes very inefficient, like batch gradient descent).
There is an optimal trade-off (in terms of computational efficiency) that may vary depending on the data distribution and the particulars of the class of function considered, as well as how computations are implemented.
Loss functions $L(y, f(\mathbf{x})) = \sum_i \ell(y_i, f(\mathbf{x_i}))$ represent the price paid for inaccuracy of predictions in classification/regression problems.
In classification this function is often the zero-one loss, that is, $ \ell(y_i, f(\mathbf{x_i}))$ is zero when $y_i = f(\mathbf{x}_i)$ and one otherwise.
This function is discontinuous with flat regions and is thus extremely hard to optimize using gradient-based methods. For this reason it is usual to consider a proxy to the loss called a surrogate loss function. For computational reasons this is usually convex function. Here we have some examples:
In regression problems, the most common loss function is the square loss function:
$$ L(y, f(\mathbf{x})) = \sum_i (y_i - f(\mathbf{x}_i))^2 $$The square loss function can be re-written and utilized for classification:
$$ L(y, f(\mathbf{x})) = \sum_i (1 - y_i f(\mathbf{x}_i))^2 $$The hinge loss function is defined as:
$$ L(y, f(\mathbf{x})) = \sum_i \mbox{max}(0, 1 - y_i f(\mathbf{x}_i)) $$The hinge loss provides a relatively tight, convex upper bound on the 0–1 Loss.
This function displays a similar convergence rate to the hinge loss function, and since it is continuous, gradient descent methods can be utilized.
$$ L(y, f(\mathbf{x})) = log(1 + exp(-y_i f(\mathbf{x}_i))) $$Cross-Entropy is a loss function that is very used for training multiclass problems. We'll focus on models that assume that classes are mutually exclusive. In this case, our labels have this form $\mathbf{y}_i =(1.0,0.0,0.0)$. If our model predicts a different distribution, say $ f(\mathbf{x}_i)=(0.4,0.1,0.5)$, then we'd like to nudge the parameters so that $f(\mathbf{x}_i)$ gets closer to $\mathbf{y}_i$.
C.Shannon showeed that if you want to send a series of messages composed of symbols from an alphabet with distribution $y$ ($y_j$ is the probability of the $j$-th symbol), then to use the smallest number of bits on average, you should assign $\log(\frac{1}{y_j})$ bits to the $j$-th symbol.
The optimal number of bits is known as entropy:
$$ H(\mathbf{y}) = \sum_j y_j \log\frac{1}{y_j} = - \sum_j y_j \log y_j$$Cross entropy is the number of bits we'll need if we encode symbols by using a wrong distribution $\hat y$:
$$ H(y, \hat y) = - \sum_j y_j \log \hat y_j $$In our case, the real distribution is $\mathbf{y}$ and the "wrong" one is $f(\mathbf{x}_i)$. So, minimizing cross entropy with respect our model parameters will result in the model that best approximates our labels if considered as a probabilistic distribution.
Cross entropy is used in combination with Softmax classifier. In order to classify $\mathbf{x}_i$ we could take the index corresponding to the max value of $f(\mathbf{x}_i)$, but Softmax gives a slightly more intuitive output (normalized class probabilities) and also has a probabilistic interpretation:
$$ P(\mathbf{y}_i = j \mid \mathbf{x_i}) = - log \left( \frac{e^{f_j(\mathbf{x_i})}}{\sum_k e^{f_k(\mathbf{x_i})} } \right) $$where $f_k$ is a linear classifier.
See "An Interactive Tutorial on Numerical Optimization": http://www.benfrederickson.com/numerical-optimization/
SGD has trouble navigating ravines, i.e. areas where the surface curves much more steeply in one dimension than in another, which are common around local optima. In these scenarios, SGD oscillates across the slopes of the ravine while only making hesitant progress along the bottom towards the local optimum.
Momentum is a method that helps accelerate SGD in the relevant direction and dampens oscillations. It does this by adding a fraction of the update vector of the past time step to the current update vector:
$$ v_t = m v_{t-1} + \alpha \nabla_w f $$$$ w = w - v_t $$The momentum $m$ is commonly set to $0.9$.
However, a ball that rolls down a hill, blindly following the slope, is highly unsatisfactory. We'd like to have a smarter ball, a ball that has a notion of where it is going so that it knows to slow down before the hill slopes up again.
Nesterov accelerated gradient (NAG) is a way to give our momentum term this kind of prescience. We know that we will use our momentum term $m v_{t-1}$ to move the parameters $w$. Computing $w - m v_{t-1}$ thus gives us an approximation of the next position of the parameters (the gradient is missing for the full update), a rough idea where our parameters are going to be. We can now effectively look ahead by calculating the gradient not w.r.t. to our current parameters $w$ but w.r.t. the approximate future position of our parameters:
$$ w_{new} = w - m v_{t-1} $$$$ v_t = m v_{t-1} + \alpha \nabla_{w_{new}} f $$$$ w = w - v_t $$All previous approaches manipulated the learning rate globally and equally for all parameters. Tuning the learning rates is an expensive process, so much work has gone into devising methods that can adaptively tune the learning rates, and even do so per parameter.
Adagrad is an algorithm for gradient-based optimization that does just this: It adapts the learning rate to the parameters, performing larger updates for infrequent and smaller updates for frequent parameters.
$$ c = c + (\nabla_w f)^2 $$$$ w = w - \frac{\alpha}{\sqrt{c}} $$RMSProp update adjusts the Adagrad method in a very simple way in an attempt to reduce its aggressive, monotonically decreasing learning rate. In particular, it uses a moving average of squared gradients instead, giving:
$$ c = \beta c + (1 - \beta)(\nabla_w f)^2 $$$$ w = w - \frac{\alpha}{\sqrt{c}} $$where $\beta$ is a decay rate that controls the size of the moving average.
(Image credit: Alec Radford)
(Image credit: Alec Radford)
In [ ]: