CS231n Convolutional Neural Networks For Visual Recognition
CS231n Convolutional Neural Networks For Visual Recognition
Course Website
This is an introductory lecture designed to introduce people from outside of Computer Vision to the Image
Classi;cation problem, and the data-driven approach. The Table of Contents:
Image Classi;cation
Nearest Neighbor Classi;er
k - Nearest Neighbor Classi;er
Validation sets for Hyperparameter tuning
Summary
Summary: Applying kNN in practice
Further Reading
Image Classification
Motivation
Motivation. In this section we will introduce the Image Classi;cation problem, which is the task of assigning
an input image one label from a ;xed set of categories. This is one of the core problems in Computer Vision
that, despite its simplicity, has a large variety of practical applications. Moreover, as we will see later in the
course, many other seemingly distinct Computer Vision tasks (such as object detection, segmentation) can be
reduced to image classi;cation.
Example
Example. For example, in the image below an image classi;cation model takes a single image and assigns
probabilities to 4 labels, {cat, dog, hat, mug}. As shown in the image, keep in mind that to a computer an image
is represented as one large 3-dimensional array of numbers. In this example, the cat image is 248 pixels wide,
400 pixels tall, and has three color channels Red,Green,Blue (or RGB for short). Therefore, the image consists of
248 x 400 x 3 numbers, or a total of 297,600 numbers. Each number is an integer that ranges from 0 (black) to
255 (white). Our task is to turn this quarter of a million numbers into a single label, such as “cat”.
Whatthecomputersees
82%cat
15%dog
imageclassification
2%hat
1%mug
The task in Image Classi6cation is to predict a single label (or a distribution over labels as shown here to indicate our
con6dence) for a given image. Images are 3-dimensional arrays of integers from 0 to 255, of size Width x Height x 3. The 3
represents the three color channels Red, Green, Blue.
Challenges
Challenges. Since this task of recognizing a visual concept (e.g. cat) is relatively trivial for a human to
perform, it is worth considering the challenges involved from the perspective of a Computer Vision algorithm.
As we present (an inexhaustive) list of challenges below, keep in mind the raw representation of images as a 3-
D array of brightness values:
Viewpoint variation
variation. A single instance of an object can be oriented in many ways with respect to the
camera.
Scale variation
variation. Visual classes often exhibit variation in their size (size in the real world, not only in
terms of their extent in the image).
Deformation
Deformation. Many objects of interest are not rigid bodies and can be deformed in extreme ways.
Occlusion
Occlusion. The objects of interest can be occluded. Sometimes only a small portion of an object (as
little as few pixels) could be visible.
Illumination conditions
conditions. The effects of illumination are drastic on the pixel level.
clutter. The objects of interest may blend into their environment, making them hard to
Background clutter
identify.
variation. The classes of interest can often be relatively broad, such as chair. There are
Intra-class variation
many different types of these objects, each with their own appearance.
A good image classi;cation model must be invariant to the cross product of all these variations, while
simultaneously retaining sensitivity to the inter-class variations.
Data-driven approach
approach. How might we go about writing an algorithm that can classify images into distinct
categories? Unlike writing an algorithm for, for example, sorting a list of numbers, it is not obvious how one
might write an algorithm for identifying cats in images. Therefore, instead of trying to specify what every one of
the categories of interest look like directly in code, the approach that we will take is not unlike one you would
take with a child: we’re going to provide the computer with many examples of each class and then develop
learning algorithms that look at these examples and learn about the visual appearance of each class. This
approach is referred to as a data-driven approach, since it relies on ;rst accumulating a training dataset of
labeled images. Here is an example of what such a dataset might look like:
FINTCE
An example training set for four visual categories. In practice we may have thousands of categories and hundreds of
thousands of images for each category.
Input: Our input consists of a set of N images, each labeled with one of K different classes. We refer to
this data as the training set.
Learning: Our task is to use the training set to learn what every one of the classes looks like. We refer to
this step as training a classi;er, or learning a model.
Evaluation: In the end, we evaluate the quality of the classi;er by asking it to predict labels for a new set
of images that it has never seen before. We will then compare the true labels of these images to the ones
predicted by the classi;er. Intuitively, we’re hoping that a lot of the predictions match up with the true
answers (which we call the ground truth).
Example image classi6cation dataset: CIFAR-10. One popular toy image classi;cation dataset is the
CIFAR-10 dataset. This dataset consists of 60,000 tiny images that are 32 pixels high and wide. Each image is
labeled with one of 10 classes (for example “airplane, automobile, bird, etc”). These 60,000 images are
partitioned into a training set of 50,000 images and a test set of 10,000 images. In the image below you can
see 10 random example images from each one of the 10 classes:
airplane
automobile
bird
cat
deer
dog
frog
horse
ship
truck
Left: Example images from the CIFAR-10 dataset. Right: 6rst column shows a few test images and next to each we show
the top 10 nearest neighbors in the training set according to pixel-wise difference.
Suppose now that we are given the CIFAR-10 training set of 50,000 images (5,000 images for every one of the
labels), and we wish to label the remaining 10,000. The nearest neighbor classi;er will take a test image,
compare it to every single one of the training images, and predict the label of the closest training image. In the
image above and on the right you can see an example result of such a procedure for 10 example test images.
Notice that in only about 3 out of 10 examples an image of the same class is retrieved, while in the other 7
examples this is not the case. For example, in the 8th row the nearest training image to the horse head is a red
car, presumably due to the strong black background. As a result, this image of a horse would in this case be
mislabeled as a car.
You may have noticed that we left unspeci;ed the details of exactly how we compare two images, which in this
case are just two blocks of 32 x 32 x 3. One of the simplest possibilities is to compare the images pixel by pixel
and add up all the differences. In other words, given two images and representing them as vectors I1 , I2 , a
reasonable choice for comparing them might be the L1 distance
distance:
Where the sum is taken over all pixels. Here is the procedure visualized:
An example of using pixel-wise differences to compare two images with L1 distance (for one color channel in this
example). Two images are subtracted elementwise and then all differences are added up to a single number. If two images
are identical the result will be zero. But if the images are very different the result will be large.
Let’s also look at how we might implement the classi;er in code. First, let’s load the CIFAR-10 data into memory
as 4 arrays: the training data/labels and the test data/labels. In the code below, Xtr (of size 50,000 x 32 x 32 x
3) holds all the images in the training set, and a corresponding 1-dimensional array Ytr (of length 50,000)
holds the training labels (from 0 to 9):
Now that we have all images stretched out as rows, here is how we could train and evaluate a classi;er:
import numpy as np
class NearestNeighbor(object):
def __init__(self):
pass
return Ypred
If you ran this code, you would see that this classi;er only achieves 38.6% on CIFAR-10. That’s more
impressive than guessing at random (which would give 10% accuracy since there are 10 classes), but nowhere
near human performance (which is estimated at about 94%) or near state-of-the-art Convolutional Neural
Networks that achieve about 95%, matching human accuracy (see the leaderboard of a recent Kaggle
competition on CIFAR-10).
The choice of distance. There are many other ways of computing distances between vectors. Another
common choice could be to instead use the L2 distance
distance, which has the geometric interpretation of
computing the euclidean distance between two vectors. The distance takes the form:
−−−−−−−−−−−
d2 (I1 , I2 ) = √∑ (I1p − I2p )2
p
In other words we would be computing the pixelwise difference as before, but this time we square all of them,
add them up and ;nally take the square root. In numpy, using the code from above we would need to only
replace a single line of code. The line that computes the distances:
Note that I included the np.sqrt call above, but in a practical nearest neighbor application we could leave out
the square root operation because square root is a monotonic function. That is, it scales the absolute sizes of
the distances but it preserves the ordering, so the nearest neighbors with or without it are identical. If you ran
the Nearest Neighbor classi;er on CIFAR-10 with this distance, you would obtain 35.4% accuracy (slightly
lower than our L1 distance result).
L1 vs. L2. It is interesting to consider differences between the two metrics. In particular, the L2 distance is
much more unforgiving than the L1 distance when it comes to differences between two vectors. That is, the L2
distance prefers many medium disagreements to one big one. L1 and L2 distances (or equivalently the L1/L2
norms of the differences between a pair of images) are the most commonly used special cases of a p-norm.
An example of the difference between Nearest Neighbor and a 5-Nearest Neighbor classi6er, using 2-dimensional points
and 3 classes (red, blue, green). The colored regions show the decision boundaries induced by the classi6er with an L2
distance. The white regions show points that are ambiguously classi6ed (i.e. class votes are tied for at least two classes).
Notice that in the case of a NN classi6er, outlier datapoints (e.g. green point in the middle of a cloud of blue points) create
small islands of likely incorrect predictions, while the 5-NN classi6er smooths over these irregularities, likely leading to
better generalization on the test data (not shown). Also note that the gray regions in the 5-NN image are caused by ties in
the votes among the nearest neighbors (e.g. 2 neighbors are red, next two neighbors are blue, last neighbor is green).
In practice, you will almost always want to use k-Nearest Neighbor. But what value of k should you use? We
turn to this problem next.
You might be tempted to suggest that we should try out many different values and see what works best. That
is a ;ne idea and that’s indeed what we will do, but this must be done very carefully. In particular, we cannot
use the test set for the purpose of tweaking hyperparameters
hyperparameters. Whenever you’re designing Machine
Learning algorithms, you should think of the test set as a very precious resource that should ideally never be
touched until one time at the very end. Otherwise, the very real danger is that you may tune your
hyperparameters to work well on the test set, but if you were to deploy your model you could see a signi;cantly
reduced performance. In practice, we would say that you over6t to the test set. Another way of looking at it is
that if you tune your hyperparameters on the test set, you are effectively using the test set as the training set,
and therefore the performance you achieve on it will be too optimistic with respect to what you might actually
observe when you deploy your model. But if you only use the test set once at end, it remains a good proxy for
measuring the generalization of your classi;er (we will see much more discussion surrounding
generalization later in the class).
Evaluate on the test set only a single time, at the very end.
Luckily, there is a correct way of tuning the hyperparameters and it does not touch the test set at all. The idea is
to split our training set in two: a slightly smaller training set, and what we call a validation set
set. Using CIFAR-10
as an example, we could for example use 49,000 of the training images for training, and leave 1,000 aside for
validation. This validation set is essentially used as a fake test set to tune the hyper-parameters.
By the end of this procedure, we could plot a graph that shows which values of k work best. We would then
stick with this value and evaluate once on the actual test set.
Split your training set into training set and a validation set. Use validation set to tune all hyperparameters. At
the end run a single time on the test set and report performance.
Cross-validation
Cross-validation. In cases where the size of your training data (and therefore also the validation data) might
be small, people sometimes use a more sophisticated technique for hyperparameter tuning called cross-
validation
validation. Working with our previous example, the idea is that instead of arbitrarily picking the ;rst 1000
datapoints to be the validation set and rest training set, you can get a better and less noisy estimate of how
well a certain value of k works by iterating over different validation sets and averaging the performance across
these. For example, in 5-fold cross-validation, we would split the training data into 5 equal folds, use 4 of them
for training, and 1 for validation. We would then iterate over which fold is the validation fold, evaluate the
performance, and ;nally average the performance across the different folds.
In practice
practice. In practice, people prefer to avoid cross-validation in favor of having a single validation split, since
cross-validation can be computationally expensive. The splits people tend to use is between 50%-90% of the
training data for training and rest for validation. However, this depends on multiple factors: For example if the
number of hyperparameters is large you may prefer to use bigger validation splits. If the number of examples in
the validation set is small (perhaps only a few hundred or so), it is safer to use cross-validation. Typical number
of folds you can see in practice would be 3-fold, 5-fold or 10-fold cross-validation.
Common data splits. A training and test set is given. The training set is split into folds (for example 5 folds here). The folds
1-4 become the training set. One fold (e.g. fold 5 here in yellow) is denoted as the Validation fold and is used to tune the
hyperparameters. Cross-validation goes a step further and iterates over the choice of which fold is the validation fold,
separately from 1-5. This would be referred to as 5-fold cross-validation. In the very end once the model is trained and all
the best hyperparameters were determined, the model is evaluated a single time on the test data (red).
It is worth considering some advantages and drawbacks of the Nearest Neighbor classi;er. Clearly, one
advantage is that it is very simple to implement and understand. Additionally, the classi;er takes no time to
train, since all that is required is to store and possibly index the training data. However, we pay that
computational cost at test time, since classifying a test example requires a comparison to every single training
example. This is backwards, since in practice we often care about the test time elciency much more than the
elciency at training time. In fact, the deep neural networks we will develop later in this class shift this tradeoff
to the other extreme: They are very expensive to train, but once the training is ;nished it is very cheap to
classify a new test example. This mode of operation is much more desirable in practice.
As an aside, the computational complexity of the Nearest Neighbor classi;er is an active area of research, and
several Approximate Nearest Neighbor (ANN) algorithms and libraries exist that can accelerate the nearest
neighbor lookup in a dataset (e.g. FLANN). These algorithms allow one to trade off the correctness of the
nearest neighbor retrieval with its space/time complexity during retrieval, and usually rely on a pre-
processing/indexing stage that involves building a kdtree, or running the k-means algorithm.
The Nearest Neighbor Classi;er may sometimes be a good choice in some settings (especially if the data is
low-dimensional), but it is rarely appropriate for use in practical image classi;cation settings. One problem is
that images are high-dimensional objects (i.e. they often contain many pixels), and distances over high-
dimensional spaces can be very counter-intuitive. The image below illustrates the point that the pixel-based L2
similarities we developed above are very different from perceptual similarities:
Pixel-based distances on high-dimensional data (and images especially) can be very unintuitive. An original image (left) and
three other images next to it that are all equally far away from it based on L2 pixel distance. Clearly, the pixel-wise distance
does not correspond at all to perceptual or semantic similarity.
Here is one more visualization to convince you that using pixel differences to compare images is inadequate.
We can use a visualization technique called t-SNE to take the CIFAR-10 images and embed them in two
dimensions so that their (local) pairwise distances are best preserved. In this visualization, images that are
shown nearby are considered to be very near according to the L2 pixelwise distance we developed above:
CIFAR-10 images embedded in two dimensions with t-SNE. Images that are nearby on this image are considered to be close
based on the L2 pixel distance. Notice the strong effect of background rather than semantic class differences. Click here
for a bigger version of this visualization.
In particular, note that images that are nearby each other are much more a function of the general color
distribution of the images, or the type of background rather than their semantic identity. For example, a dog can
be seen very near a frog since both happen to be on white background. Ideally we would like images of all of
the 10 classes to form their own clusters, so that images of the same class are nearby to each other regardless
of irrelevant characteristics and variations (such as the background). However, to get this property we will have
to go beyond raw pixels.
Summary
In summary:
In next lectures we will embark on addressing these challenges and eventually arrive at solutions that give 90%
accuracies, allow us to completely discard the training set once learning is complete, and they will allow us to
evaluate a test image in less than a millisecond.
1. Preprocess your data: Normalize the features in your data (e.g. one pixel in images) to have zero mean
and unit variance. We will cover this in more detail in later sections, and chose not to cover data
normalization in this section because pixels in images are usually homogeneous and do not exhibit
widely different distributions, alleviating the need for data normalization.
2. If your data is very high-dimensional, consider using a dimensionality reduction technique such as PCA
(wiki ref, CS229ref, blog ref), NCA (wiki ref, blog ref), or even Random Projections.
3. Split your training data randomly into train/val splits. As a rule of thumb, between 70-90% of your data
usually goes to the train split. This setting depends on how many hyperparameters you have and how
much of an inmuence you expect them to have. If there are many hyperparameters to estimate, you
should err on the side of having larger validation set to estimate them effectively. If you are concerned
about the size of your validation data, it is best to split the training data into folds and perform cross-
validation. If you can afford the computational budget it is always safer to go with cross-validation (the
more folds the better, but more expensive).
4. Train and evaluate the kNN classi;er on the validation data (for all folds, if doing cross-validation) for
many choices of k (e.g. the more the better) and across different distance types (L1 and L2 are good
candidates)
5. If your kNN classi;er is running too long, consider using an Approximate Nearest Neighbor library (e.g.
FLANN) to accelerate the retrieval (at cost of some accuracy).
n. Take note of the hyperparameters that gave the best results. There is a question of whether you should
use the full training set with the best hyperparameters, since the optimal hyperparameters might change
if you were to fold the validation data into your training set (since the size of the data would be larger). In
practice it is cleaner to not use the validation data in the ;nal classi;er and consider it to be burned on
estimating the hyperparameters. Evaluate the best model on the test set. Report the test set accuracy
and declare the result to be the performance of the kNN classi;er on your data.
Further Reading
Here are some (optional) links you may ;nd interesting for further reading:
A Few Useful Things to Know about Machine Learning, where especially section 6 is related but the
whole paper is a warmly recommended reading.
Recognizing and Learning Object Categories, a short course of object categorization at ICCV 2005.
cs231n
cs231n
Back to Top