Poor starting points in machine learning

Mark Tygert

Facebook Artificial Intelligence Research

1 Facebook Way, Menlo Park, CA 94025

[email protected]

January 13, 2016

Abstract
Poor (even random) starting points for learning/training/optimization are common
in machine learning. In many settings, the method of Robbins and Monro (online
stochastic gradient descent) is known to be optimal for good starting points, but may
not be optimal for poor starting points indeed, for poor starting points Nesterov
acceleration can help during the initial iterations, even though Nesterov methods not
designed for stochastic approximation could hurt during later iterations. A good option
is to roll off Nesterov acceleration for later iterations. The common practice of training
with nontrivial minibatches enhances the advantage of Nesterov acceleration.

1 Introduction
The scheme of Robbins and Monro (online stochastic gradient descent) has long been known
to be optimal for stochastic approximation/optimization . . . provided that the starting point
for the iterations is good. Yet poor starting points are common in machine learning, so
higher-order methods (such as Nesterov acceleration) may help during the early iterations
(even though higher-order methods generally hurt during later iterations), as explained
below. Below, we elaborate on Section 2 of Sutskever et al. (2013), giving some more
rigorous mathematical details, somewhat similar to those given by Defossez and Bach (2015)
and Flammarion and Bach (2015). In particular, we discuss the key role of the noise level
of estimates of the objective function, and stress that a large minibatch size makes higher-
order methods more effective for more iterations. Our presentation is purely expository;
we disavow any claim we might make to originality in fact, our observations should be
obvious to experts in stochastic approximation. We opt for concision over full mathematical
rigor, and assume that the reader is familiar with the subjects reviewed by Bach (2013).
The remainder of the present paper has the following structure: Section 2 sets math-
ematical notation for the stochastic approximation/optimization considered in the sequel.
Section 3 elaborates a principle for accelerating the optimization. Section 4 discusses mini-
batching. Section 5 supplements the numerical examples of Bach (2013) and Sutskever et al.
(2013). Section 6 reiterates the above, concluding the paper.

1
2 A mathematical formulation
Machine learning commonly involves a cost/loss function c and a single random sample X
of test data; we then want to select parameters minimizing the expected value

e() = E[c(, X)] (1)

(e is known as the error or risk specifically, the test error, as opposed to training
error). The cost c is generally a nonnegative scalar-valued function, whereas and X
can be vectors of real or complex numbers. In supervised learning, X is a vector which
contains both input or feature entries and corresponding output or target entries
for example, for a regression the input entries can be regressors and the outputs can
be regressands; for a classification the outputs can be the labels of the correct classes. For
training, we have no access to X (nor to e) directly, but can only access independent and
identically distributed (i.i.d.) samples from the probability distribution p from which X
arises. Of course, e() depends only on p, not on any particular sample drawn from p.
To finish setting our notational conventions, we define the deviation or dispersion

d(, X) = c(, X) e(), (2)

yielding the decomposition

c(, X) = e() + d(, X). (3)
Combining (2) and (1) yields that

E[d(, X)] = 0 (4)

for all , that is, the mean of d(, X) is 0 (whereas the mean of c(, X) is e()).

3 Accelerated learning
For definiteness, we limit consideration to algorithms which may access only values of c(, x)
and its first-order derivatives/gradient with respect to . We assume that the cost c is
calibrated (shifted by an appropriate constant) such that the absolute magnitude of c is a
good absolute gauge of the quality of (lower cost is better, of course). As reviewed, for
example, by Bach (2013), the algorithm of Robbins and Monro (online stochastic gradient
descent) minimizes the test error e in (1) using a number of samples (from the probability
distribution p) that is optimal optimal to within a constant factor which depends on the
starting point for . Despite this kind of optimality, if e happens to start much larger than
d = c e, then a higher-order method (such as the Nesterov acceleration discussed by Bach
(2013) and others) can dramatically reduce the constant. We say that is poor to mean
that e() is much larger than d(, X). If is poor, that is,

e()  |d(, X)|, (5)

then (3) yields that

c(, X) = e() + d(, X) e() (6)

2
is effectively deterministic, and so Nesterov and other methods from Bach (2013) could
accelerate the optimization until the updated is no longer poor.
Clearly, being poor, that is, e() being much larger than d(, X), simply means that
c(, x) = e() + d(, x) depends mostly just on the probability distribution p underlying X,
but not so much on any particular sample x from p. (For supervised learning, the probability
distribution p is joint over the inputs and outputs, encoding the relation between the inputs
and their respective outputs.) If is poor, that is, e() is much larger than d(, X), so that
c(, X) e() is mostly deterministic, then Nesterov methods can be helpful.
Of course, optimization performance may suffer from applying a higher-order method
to a rough, random objective (say, to a random variable c(, X) estimating e() when the
standard deviation of c(, X) is high relative to e()). For instance, Nesterov methods
essentially sum across several iterations; if the values being summed were stochastically
independent, then the summation would actually increase the variance. In accordance with
the optimality of the method of Robbins and Monro (online stochastic gradient descent),
Nesterov acceleration can be effective only when the estimates of the objective function are
dominantly deterministic, with the objective function being the test error e from (1).
In practice, we can apply higher-order methods during the initial iterations when the
starting point is poor, gradually transitioning to the original method of Robbins and Monro
(stochastic gradient descent) as the test error approaches its lower limit, that is, as stochastic
variations in estimates of the objective function and its derivatives become important.

4 Minibatches
Some practical considerations detailed, for example, by LeCun et al. (1998) are the following:
Many modern microprocessor architectures can leverage the parallelism inherent in batch
processing. With such parallel (and partly parallel) processors, minimizing the number
of samples drawn from the probability distribution underlying the data may not be the
most efficient possibility. Rather than updating the parameters being learned for random
individual samples as in online stochastic gradient descent, the common current practice is
to draw all at once a number of samples a collection of samples known as a minibatch
and then update the parameters simultaneously for all these samples before proceeding
to the next minibatch. With sufficient parallelism, processing an entire minibatch may take
little longer than processing a single sample.
Averaging the estimates of the objective function (that is, of the test error e) and its
derivatives over all samples in a minibatch yields estimates with smaller standard deviations,
effectively making more parameter values be poor, so that Nesterov acceleration is more
effective for more iterations. Moreover, minibatches provide (essentially for free) estimates
of the standard deviations of the estimates of e. The number of initial iterations for which
Nesterov methods can accelerate the optimization can be made arbitrarily large by setting
the size of the minibatches arbitrarily large. That said, smaller minibatch sizes typically
require fewer samples in total to approach full convergence (albeit typically with more
iterations, processing one minibatch per iteration). Also, after sufficiently many iterations,
stochastic variations in estimates of the objective function and its derivatives can become
important, and continuing to use Nesterov acceleration in these later iterations would be
counterproductive. Again, we recommend applying higher-order methods during the initial

3
iterations when the starting point is poor, noting that larger minibatches make the higher-
order methods advantageous for more iterations before the point requiring turning off the
Nesterov acceleration.

5 Numerical experiments
The present section supplements the examples of Bach (2013) and Sutskever et al. (2013)
with a few more experiments indicating that a higher-order method namely, momentum,
a form of Nesterov acceleration can help a bit in training convolutional networks. All
training reported is via the method of Robbins and Monro (stochastic gradient descent),
with minibatches first of size 1 sample per iteration (as in the original, online algorithm)
and then of size 100 samples per iteration (size 100 is among the most common choices
used in practice on modern processors which can leverage the parallelism inherent in batch
processing). Section 4 and the third-to-last paragraph of the present section describe the
minibatching in more detail. Figures 1 and 3 report the accuracies for various configurations
of momentum and learning rates; momentum appears to accelerate training somewhat,
especially with the larger size of minibatches.
Each iteration subtracts from the parameters being learned a multiple of a stored vector,
where the multiple is the learning rate, and the stored vector is the estimated gradient
plus the stored vector from the previous iteration, with the latter stored vector scaled by
the amount of momentum:
(i)
(i) = (i1) l.r . v (i) , (7)
with
k  
(i) (i) 1 X c 
v = mom. v (i1) + , x (i,j)
 (i1) ,
 (8)
k =
j=1

where i is the index of the iteration, (i) is the updated vector of parameters, (i1) is
the previous vector of parameters, v (i) is the updated stored auxiliary vector, v (i1) is the
(i) (i)
previous stored auxiliary vector, l.r . is the learning rate, mom. is the amount of momentum,
x(i,j) is the jth training sample in the ith minibatch, k is the size of the minibatch, and
c/ is the gradient of the cost c with respect to the parameters .
Following LeCun et al. (1998) exactly as done by Chintala et al. (2015), the architecture
for generating the feature activations is a convolutional network (convnet) consisting of a
series of stages, with each stage feeding its output into the next. The last stage has the form
of a multinomial logistic regression, applying a linear transformation to its inputs, followed
by the softmax detailed by LeCun et al. (1998), thus producing a probability distribution
over the classes in the classification. The cost/loss is the negative of the natural logarithm
of the probability so assigned to the correct class. Each stage before the last convolves each
image from its input against several learned convolutional kernels, summing together the
convolved images from all the inputs into several output images, then takes the absolute
value of each pixel of each resulting image, and finally averages over each patch in a partition
of each image into a grid of 2 2 patches. All convolutions are complex valued and produce
pixels only where the original images cover all necessary inputs (that is, a convolution
reduces each dimension of the image by one less than the size of the convolutional kernel).

4
We subtract the mean of the pixel values from each input image before processing with
the convnet, and we append an additional feature activation feeding into the softmax to
those obtained from the convnet, namely the standard deviation of the set of values of the
pixels in the image.
The data is a subset of the 2012 ImageNet set of Russakovsky et al. (2015), retaining
10 classes of images, representing each class by 100 samples in a training set and 50 per
class in a testing set. Restricting to this subset facilitated more extensive experimentation
(optimizing hyperparameters more extensively, for example). The images are full color, with
three color channels. We neither augmented the input data nor regularized the cost/loss
functions. We used the Torch7 platform http://torch.ch for all computations.
Table 1 details the convnet architecture we tested. Stage specifies the positions of
the indicated layers in the convnet. Input channels specifies the number of images input
to the given stage for each sample from the data. Output channels specifies the number
of images output from the given stage. Each input image is convolved against a separate,
learned convolutional kernel for each output image (with the results of all these convolutions
summed together for each output image). Kernel size specifies the size of the square grid
of pixels used in the convolutions. Input channel size specifies the size of the square
grid of pixels constituting each input image. Output channel size specifies the size of
the square grid of pixels constituting each output image. The feature activations that the
convnet produces feed into a linear transformation followed by a softmax, as detailed
by LeCun et al. (1998).
For the minibatch size 100, rather than updating the parameters being learned for
randomly selected individual images from the training set as in online stochastic gradient
descent, we instead do the following: we randomly permute the training set and partition
this permuted set of images into subsets of 100, updating the parameters simultaneously
for all 100 images constituting each of the subsets (known as minibatches), processing
the series of minibatches in series. Each sweep through the entire training set is known as
an epoch. LeCun et al. (1998), among others, made the above terminology the standard
for training convnets. The horizontal axes in the figures count the number of epochs.
Figures 1 and 2 present the results for minibatches of size 100 samples per iteration;
Figure 3 presents the results for minibatches of size 1 sample per iteration. Average
precision is the fraction of all classifications which are correct, choosing only one class for
each input sample image from the test set. Error on the test set is the average over all
samples in the test set of the negative of the natural logarithm of the probability assigned
to the correct class (assigned by the softmax). Coefficient of variation is an estimate
of the standard deviation of c(, X) (from (1)) divided by the mean of c(, X), that is, an
estimate of the standard deviation of d(, X) (from (2)) divided by e() (from (1) and (2)).
Rolling off momentum as the coefficient of variation increases can be a good idea.
Please beware that these experiments are far from definitive, and even here the gains
from using momentum seem to be marginal. Even so, as Section 4 discusses, minibatching
effectively reduces the standard deviations of estimates of the objective function e, making
Nesterov acceleration more effective for more iterations.

5
6 Conclusion
Though in many settings the method of Robbins and Monro (stochastic gradient descent)
is optimal for good starting points, higher-order methods (such as momentum and Nesterov
acceleration) can help during early iterations of the optimization when the parameters
being optimized are poor in the sense discussed above. The opportunity for accelerating
the optimization is clear theoretically and apparently observable via numerical experiments.
Minibatching makes the higher-order methods advantageous for more iterations. That said,
higher order and higher accuracy need not be the same, as higher order guarantees only
that accuracy increase at a faster rate higher order guarantees higher accuracy only after
enough iterations.

Acknowledgements
We would like to thank Francis Bach, Leon Bottou, Yann Dauphin, Piotr Dollar, Yann
LeCun, MarcAurelio Ranzato, Arthur Szlam, and Rachel Ward.

A A simple analytical example

This appendix works through a very simple one-dimensional example.
Consider the cost c as a function of a real scalar parameter and sample of data x,
defined via
c(, x) = ( x)2 . (9)
Suppose that X is the (Rademacher) random variable taking the value 1 with probability
1/2 and the value 1 with probability 1/2. Then, the coefficient of variation of c(, X),
that is, the standard deviation () of c(, X) divided by the mean e() of c(, X), is

() 2||
= 2 : (10)
e() +1

(proof ) X taking the value 1 with probability 1/2 and the value 1 with probability 1/2
yields that
|X| = 1 (11)
and that the expected value of X is 0,

E[X] = 0. (12)

Combining (9) and (11) then yields

c(, X) = 2 + 1 2X. (13)

Moreover, together with the definitions (1) and (2), combining (13) and (12) yields

e() = E[c(, X)] = 2 + 1 (14)

6
and
d(, X) = c(, X) e() = 2X. (15)
The mean of c(, X) is this e() in (14). The standard deviation of c(, X) is the same as
the standard deviation of d(, X) in (15), which is just

() = 2||, (16)

due to (15) combined with X taking the value 1 with probability 1/2 and the value 1 with
probability 1/2. Combining (14) and (16) yields (10), as desired. 

The secant method (a higher-order method) consists of the iterations updating (i2)
and (i1) to (i) via
c  (i1) (i1) 
(i) = (i1) ,x

    c   c  
(i1) (i2) (i1) (i1) (i2) (i2)
,x ,x , (17)

where x(i2) , x(i1) , and x(i) are independent realizations (random samples) of X and the
derivative of c(, x) from (9) with respect to is
c
= 2( x); (18)

combining (17) and (18) and simplifying yields
  
(i) (i1) (i2) (i2) (i1)
= x x (i1) x(i1) (i2) + x(i2) . (19)

In the limit that both x(i1) and x(i2) be 0, (19) shows that (i) is also 0 (provided that
(i1) and (i2) are distinct), that is, the secant method finds the optimal value for in a
single iteration in such a limit. In fact, (14) makes clear that the optimal value for is 0,
while combining (19) and (11) yields that
  
(i) (i1) (i2)
| | | | + | | |(i1) (i2) | 2 , (20)

so that |(i) | is likely to be reasonably small (around 1 or so) even if |(i1) | or |(i2) | (or
both) are very large and random and |X| = 1.
However, these iterations generally fail to converge to any value much smaller than unit
magnitude: with probability 1/2, indeed, x(i2) = x(i1) = 1 or x(i2) = x(i1) = 1; hence
(19) yields that |(i) | = 1 with probability at least 1/2, assuming that (i2) and (i1) are
distinct (while in the degenerate case that (i2) and (i1) take exactly the same value, (i)
also takes that same value).
All in all, a sensible strategy is to start with a higher-order method (such as the secant
method) when || is large, and transition to the asymptotically optimal method of Robbins
and Monro (stochastic gradient descent) as || becomes of roughly unit magnitude. The
transition can be based on || or on estimates of the coefficient of variation due to (10),
the coefficient of variation is essentially inversely proportional to || when || is large.

7
Figure 1: Test accuracies for various settings of momentum (mom.) and learning rates (l.r.)
[minibatch size 100]

higher average-precision is be0er

0.6
mom.=.9, l.r.=8,
0.5 when epoch<20;
average precision

mom.=0, l.r.=.02,
0.4 otherwise

0.3 mom.=.9, l.r.=8,

(a)
when epoch<20;
0.2 mom.=.9, l.r.=.02,
otherwise
0.1
mom.=0, l.r.=3,
0 when epoch<30;
0 10 20 30 40 50 mom.=0, l.r.=.9,
otherwise
epoch

lower error on the test set is be/er

2.5
mom.=.9, l.r.=8,
2 when epoch<20;
error on the test set

mom.=0, l.r.=.02,
otherwise
1.5
mom.=.9, l.r.=8,
(b)
1 when epoch<20;
mom.=.9, l.r.=.02,
otherwise
0.5
mom.=0, l.r.=3,
0 when epoch<30;
0 10 20 30 40 50 mom.=0, l.r.=.9,
otherwise
epoch

8
Figure 2: Estimated coefficients of variation (CVs) of the cost c(, X) in (1) and (3), using
all samples in each minibatch to construct estimates of the CVs (the estimates of the CVs
are themselves subject to stochastic variations; the distribution of the plotted points within
an epoch provides an indication of the probability distribution of the estimates of the CVs)
[minibatch size 100]

1.2
the cost's std. dev. divided by its mean
coecient of varia-on of the cost c(, X)

mom.=.9, l.r.=8,
1
when epoch<20;
mom.=0, l.r.=.02,
0.8 otherwise

0.6 mom.=.9, l.r.=8,

when epoch<20;
mom.=.9, l.r.=.02,
0.4
otherwise

0.2 mom.=0, l.r.=3,

when epoch<30;
0 mom.=0, l.r.=.9,
0 10 20 30 40 50 otherwise
epoch

Table 1: Architecture of the convolutional network

input output kernel input output

stage channels channels size channel size channel size
first 3 16 55 128 128 62 62
second 16 64 33 62 62 30 30
third 64 256 33 30 30 14 14
fourth 256 256 33 14 14 66

9
Figure 3: Test accuracies for various settings of momentum (mom.) and learning rates (l.r.)
[minibatch size 1]

higher average-precision is be0er

0.6
mom.=.6, l.r.=.002,
0.5 when epoch<8;
average precision

mom.=0, l.r.=.0002,
0.4 otherwise

0.3 mom.=.6, l.r.=.002,

(a)
when epoch<8;
0.2 mom.=.6, l.r.=.0002,
otherwise
0.1
mom.=0, l.r.=.002,
0 when epoch<10;
0 5 10 15 20 mom.=0, l.r.=.0002,
otherwise
epoch

lower error on the test set is be/er

2.5
mom.=.6, l.r.=.002,
2 when epoch<8;
error on the test set

mom.=0, l.r.=.0002,
otherwise
1.5
mom.=.6, l.r.=.002,
(b)
1 when epoch<8;
mom.=.6, l.r.=.0002,
otherwise
0.5
mom.=0, l.r.=.002,
0 when epoch<10;
0 5 10 15 20 mom.=0, l.r.=.0002,
otherwise
epoch

10
References
Bach, F. (2013). Stochastic gradient methods for machine learning. Technical report,
INRIA-ENS, Paris, France. http://www.di.ens.fr/~fbach/fbach_sgd_online.pdf.

Chintala, S., Ranzato, M., Szlam, A., Tian, Y., Tygert, M., and Zaremba, W. (2015).
Scale-invariant learning and convolutional networks. Technical Report 1506.08230, arXiv.
http://arxiv.org/abs/1506.08230.

Defossez, A. and Bach, F. (2015). Averaged least-mean-squares: bias-variance trade-offs

and optimal sampling distributions. In Proc. 18th Internat. Conf. Artificial Intelligence
and Statistics, volume 38, pages 205213. http://jmlr.org/proceedings/papers/v38/
defossez15.pdf.

Flammarion, N. and Bach, F. (2015). From averaging to acceleration, there is only a step-
size. Technical Report 1504.01577, arXiv. http://arxiv.org/abs/1504.01577.

LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998). Gradient-based learning ap-
plied to document recognition. Proc. IEEE, 86(11):22782324. http://yann.lecun.
com/exdb/publis/pdf/lecun-98.pdf.

Russakovsky, O., Deng, J., Su, H., Kruse, J., Satheesh, S., Ma, S., Huang, Z., Karpathy,
A., Khosla, A., Bernstein, M., Berg, A. C., and Fei-Fei, L. (2015). ImageNet large scale
visual recognition challenge. Technical Report 1409.0575v3, arXiv. http://arxiv.org/
abs/1409.0575.

Sutskever, I., Martens, J., Dahl, G., and Hinton, G. (2013). On the importance of ini-
tialization and momentum in deep learning. In Proc. 30th International Conference
on Machine Learning, volume 28, pages 11391147. J. Machine Learning Research.
http://www.jmlr.org/proceedings/papers/v28/sutskever13.pdf.

11