What are you looking at?

Screen Shot 2018-05-06 at 18.48.40.png

In a recent blog, I described an experiment to train a deep neural network to distinguish between photographs of Vincenzo Nibali and Alejandro Valverde, using a very small data set of images. In the conclusion, I suggested that the network was probably basing its decisions more on the colours of the riders’ kit rather than on facial recognition. This article investigates what the network was actually “looking at”, in order to understand better how it was making decisions.

The issues of accountability and bias were among the topics discussed at the last NIPS conference. As machine learning algorithms are adopted across industry, it is important for companies to be able to explain how conclusions are reached. In many instances, it is not acceptable simply to rely on an impenetrable black box. AI researchers and developers need to be able to explain what is going on inside their models, in order to justify decisions taken. In doing so, some worrying instances of bias have been revealed in the selection of data used to train the algorithms.

I went back to my rider recognition model and used an approach called “Class Activation Maps” to identify which parts of the images accounted for the network’s choice of rider. Making use of the code provided in lesson 7 of the course offered by fast.ai, I took advantage of my existing small set of training, validation and test images of the two famous cyclists. Starting with a pre-trained version of ResNet34, the idea was to replace the last two layers with four new ones, the crucial one being a convolutional layer with two outputs, matching the number of cyclists in the classification task. The two outputs of this layer were 7×7 matrix representations of the relevant image.

The final predictions of the model came from a softmax of a flattened average pooling of these 7×7 representations. The softmax output gave the probabilities of Nibali and Valverde respectively. Since there was no learning beyond the final convolution, the activations of the two 7×7 matrices represented the “Nibali-ness” and “Valverde-ness” of the image. This could be displayed as a heat map on top of the image.

Examples are shown below for the validation set of 10 images of Nibali followed by 10 of Valverde. The yellow patch of the heat map highlights the part of the image that led to the prediction displayed above each image. Nine out of ten were correct for Nibali and six for Valverde.

Screen Shot 2018-05-06 at 18.10.00.png
Class Activation Maps applied to the validation set

The heat maps were very helpful in understanding the model’s decision making process. It seemed that for Nibali, his face and helmet were important, with some attention paid to the upper part of his blue Astana kit. In contrast, the network did a very good job at identifying the M on Valverde’s Moviestar kit. It was interesting to note that the network succeeded in spotting that Nibali was wearing a Specialized helmet whereas Valverde had a Catlike design. Three errors arose in the photos of his face, which was mistaken for Nibali’s. In fact, any picture of a face led to a prediction of Nibali, as demonstrated by the cropped image below that was used for training.

Screen Shot 2018-05-06 at 18.21.58

Why should that be? Looking back at the training set, it turned out that, by chance, there were far more mugshots of Nibali, while there were more photos of Valverde riding his bike, with his face obscured by sunglasses. This was an example of unintentional bias in the training data, providing a very useful lesson.

The final set of pictures shows the predictions made on the out-of-sample test set. All the predictions are correct, except the first one, where the model failed to spot the green M on Valverde’s chest and mistook the blurred background for Nibali. Otherwise the results confirmed that the network looked at Nibali’s face, the rider’s helmet or Valverde’s kit. It also remembered seeing an image of Nibali holding the Giro trophy in the training set.

Screen Shot 2018-05-06 at 18.34.38.png
Class Activation Maps applied to the test set

In conclusion, Class Activation Maps provide a useful way of visualising the activations of hidden laters in a deep neural network. This can go some way to accounting for the decisions that appear in the output. The approach can also help identify unintentional bias in the training set.

Which team is that?

Screen Shot 2018-04-11 at 11.18.09

My last blog explored the effectiveness of deep learning in spotting the difference between Vincenzo Nibali and Alejandro Valverde. Since the faces of the riders were obscured in many of the photos, it is likely that the neural network was basing its evaluations largely on the colours of their team kit. A natural next challenge is to identify a rider’s team from a photograph. This task parallels the approach to the kaggle dog breed competition used in lesson 2 of the fast.ai course on deep learning.

Eighteen World Tour teams are competing this year. So the first step was to trawl the Internet for images, ideally of riders in this year’s kit. As before, I used an automated downloader, but this posed a number of problems. For example, searching for “Astana” brings up photographs of the capital of Kazakhstan. So I narrowed things down by searching for  “Astana 2018 cycling team”. After eliminating very small images, I ended up with a total of about 9,700 images, but these still included a certain amount of junk that I did have the time to weed out, such as photos of footballers or motorcycles in the “Sky Racing Team”,.

The following small sample of training images is generally OK, though it includes images of Scott bikes rather than Mitchelton-Scott riders and  a picture of  Sunweb’s Wilco Kelderman labelled as FDJ. However, with around 500-700 images of each team, I pressed on, noting that, for some reason, there were only 166 of Moviestar and these included the old style kit.

Screen Shot 2018-04-11 at 10.18.54.png
Small sample of training images

For training on this multiple classification problem, I adopted a slightly more sophisticated approach than before. Taking a pre-trained Resnet50 model, I performed some initial fine-tuning, on images rescaled to 224×224. I settled on an optimal learning rate of 1e-3 for the final layer, while allowing some training of lower layers at much lower rates. With a view to improving generalisation, I opted to augment the training set with random changes, such as small shifts in four directions, zooming in up to 10%, adjusting lighting and left-right flips. After initial training, accuracy was 52.6% on the validation set. This was encouraging, given that random guesses would have achieved a rate of 1 in 18 or 5.6%.

Taking a pro tip from fast.ai, training proceeded with the images at a higher resolution of 299×299. The idea is to prevent overfitting during the early stages, but to improve the model later on by providing more data for each image. This raised the accuracy to 58.3% on the validation set. This figure was obtained using a trick called “test time augmentation”, where each final prediction is based on the average prediction of five different “augmented” versions of the image in question.

Given the noisy nature of some of the images used for training, I was pleased with this result, but the acid test was to evaluate performance on unseen images. So I created a test set of two images of a lead rider from each squad and asked the model to identify the team. These are the results.

75 percent right.png
75% accuracy on the test set

The trained Resnet50 correctly identified the teams of 27 out of 36 images. Interestingly, there were no predictions of MovieStar or Sky. This could be partly due to the underrepresentation of MovieStar in the training set. Froome was mistaken for AG2R and Astana, in column 7, rows 2 and 3. In the first image, his 2018 Sky kit was quite similar to Bardet’s to the left and in the second image the sky did appear to be Astana blue! It is not entirely obvious why Nibali was mistaken for Sunweb and Astana, in the top and bottom rows. However, the huge majority of predictions were correct. An overall  success rate of 75% based on an afternoon’s work was pretty amazing.

The results could certainly be improved by cleaning up the training data, but this raises an intriguing question about the efficacy of artificial intelligence. Taking a step back, I used Bing’s algorithms to find images of cycling teams in order to train an algorithm to identify cycling teams. In effect, I was training my network to reverse-engineer Bing’s search algorithm, rather than my actual objective of identifying cycling teams. If an Internet search for FDJ pulls up an image of Wilco Kelderman, my network would be inclined to suggest that he rides for the French team.

In conclusion, for this particular approach to reach or exceed human performance, expert human input is required to provide a reliable training set. This is why this experiment achieved 75%, whereas the top submissions on the dog breeds leaderboard show near perfect performance.

Valverde or Nibali?

Alejandro Valverde has kicked off the 2018 season with an impressive series of wins. Meanwhile Vincenzo Nibali delighted the tifosi with his victory in Milan San Remo. It is pretty easy to tell these two riders apart in the pictures above, but could computer distinguish between them?

Following up on my earlier blogs about neural networks, I have been taking a look at the updated version of fast.ai’s course on deep learning. With the field advancing at a rapid pace, this provides a good way to staying up to date with the state of the art. For example, there are now a couple of cheaper alternatives to AWS for accessing high powered GPUs, offered by Paperspace and Crestle. The latest fast.ai libraries include many new tools that work extremely well in practice.

There’s a view that deep learning requires hours of training on high-powered supercomputers, using thousands (or millions) of labelled examples, in order to learn to perform computer vision tasks. However, newer architectures, such as ResNet, are able to run on much smaller data sets. In order to test this, I used an image downloader to grab photos of Nibali and Valverde and manually selected about 55 decent pictures of each one.

I divided the images into a training set with about 40 images of each rider, a validation set with 10 of each and a test set containing the rest. Nibali appears in a range of different coloured jerseys, though the Astana blue is often present. Valverde is mainly wearing the old dark blue Movistar kit with a green M. There were more close-up shots of Nibali’s face than Valverde.

Screen Shot 2018-04-03 at 18.30.08.png

I was able to fine-tune a pre-trained ResNet neural network to this task, using some of the techniques from the fast.ai tool box, each designed to improve generalisation. The first trick was to augment the training set by performing minor transformations of the images at random, such as taking a mirror image, shifting left or right and zooming in a bit. The second set of tricks varied the rate of learning as the algorithm iterated repeatedly through the training set. A final useful technique created a set of variants of each test image and took the average of the predictions. Everything ran at lightning speed on a Paperspace GPU. After a run time of just a few minutes, the ResNet was able to  score 17 out of 20 on the following validation set.

Screen Shot 2018-04-03 at 18.49.27.png

The confusion matrix shows that the model correctly identified all the Nibali images, but it was wrong on three pictures of Valverde. The first incorrect image (below) shows Valverde in the red leader’s jersey of the Tour of Murcia, which is not dissimilar to Nibali’s new Bahrain Merida kit, though he was wearing red in two of his training images. In the second instance, the network was fooled by the change in colour of Moviestar’s kit, which had become rather similar to Astana’s light blue. The figure of 0.41 above the close-up image indicates that the model assigned only a 41% probability that the image was Valverde. It probably fell below the critical 50% level, in spite of the blue/green colours, because there were were far more close-up shots of Nibali than Valverde in the training set.

Overall of 17 out of 20 on the validation set is impressive. However, the network had access to the validation set during training, so this result is “in sample”. A proper  “out of sample” evaluation of the model’s ability made use the following ten images, comprising the test set that was kept aside.

Screen Shot 2018-04-03 at 21.21.59

Amazingly, the model correctly identified 9 out of the 10 pictures it had not seen before. The only error was the Valverde selfie shown in the final image. In order to work better in practice, the training set would need to include more examples of the riders’ 2018 kit. A variant of the problem would be to identify the team rather than the rider. The same network can be trained for multiple classes rather than just two.

This experiment shows that it is pretty straightforward to run state of the art image recognition tools remotely on a GPU somewhere in the cloud and come up with pretty impressive results, even with a small data set.

The next blog describes how to identify a rider’s team.



Froome versus Dumoulin

Screen Shot 2017-10-27 at 19.04.21Many commentators have been licking their lips at the prospect of head-to-head combat between Chris Froome and Tom Dumoulin at next year’s Tour de France. It is hard to make a comparison based on their results in 2017, because they managed to avoid racing each other over the entire season of UCI World Tour races, meeting only in the World Championship Individual Time Trial, where the Dutchman was victorious. But it is intriguing to ask how Dumoulin might have done in the Tour de France and the Vuelta or, indeed, how Froome might have fared in the Giro.

Inspiration for addressing these hypothetical questions comes from an unexpected source. In 2009 Netflix awarded a $1million prize to a team that improved the company’s technique for making film recommendations to its users, based on the star ratings assigned by viewers. The successful algorithm exploited the fact that viewers may enjoy the films that are highly rated by other users who have generally agreed on the ratings of the films they have seen in common. Initial approaches sought to classify films into genres or those starring particular actors, in the hope of grouping together viewers into similar categories. However, it turned out to be very difficult to identify which features of a film are important. An alternative is simply to let the computer crunch the data and identify  the key features for itself. A method called Collaborative Filtering became one of the most popular employed for recommender systems.

Our cycling problem shares certain characteristics with the Netflix challenge: instead of users, films and ratings, we have riders, races and results. Riders enter a selection of races over the season, preferring those where they hope to do well. Similar riders, for example sprinters, tend to finish high in the results of races where other sprinters also do well. Collaborative filtering should be able to exploit the fact that climbers, sprinters or TTers tend to finish close to each other, across a range of races.

This year’s UCI World Tour concluded with the Tour of Guangxi, completing the data set of results for 2017. After excluding team time trials, 883 riders entered 174 races, resulting in 26,966 finishers. Most races have up to 200 participants , so if you imagine a huge table with all the racers down the rows and all the races across the columns, the resulting matrix is “sparse” in the sense that there are lots of missing values for the riders who were not in a particular race. Collaborative Filtering aims to fill in the spaces, i.e. to estimate the position of a rider who did not enter a specific race. This is exactly what we would like to do for the Grand Tours.

It took a couple of minutes to fit a matrix factorisation Collaborative Filtering model, using keras, on my MacBook Pro. Some experimenting suggested that I needed about 50 hidden factors plus a bias to come up with a reasonable fit for this data set. Taking at random the Milan San Remo one day stage race, it did a fairly good job of predicting the top ten riders for this long, hilly race with a flat finish.

 Model fit (prediction) Rider Actual result
1 Peter_Sagan 2
2 Alexander_Kristoff 4
3 Michael_Matthews 12
4 Edvald_Boasson_Hagen 19
5 Sonny_Colbrelli 13
6 Michal_Kwiatkowski 1
7 John_Degenkolb 7
8 nacer_Bouhanni 8
9 Julian_Alaphilippe 3
10 Diego_Ulissi 40

The following figure visualises the primary factors the model derived for classifying the best riders. Sprinters are in the lower part of chart, with climbers towards the top and allrounders in the middle. Those with a lot of wins are towards the left.

Screen Shot 2017-10-27 at 19.26.17

Now we come to the interesting part: how would Tom Dumoulin and Chris Froome have compared in the other’s Grand Tours? Note that this model takes account of the results of all the riders in all the races, so it should be capable of detecting the benefit of being part of a strong team.

Tour de France

The model suggested that Tom Dumoulin would have beaten Chris Froome in stages 1(TT), 2, 5, 6, 10 and 21, but the yellow jersey winner would have been stronger in the mountains and won overall.

Giro d’Italia

The model suggested that Chris Froome would have been ahead in the majority of stages, leaving stages 4, 5, 6, 9,  10(TT), 14 and 21(TT) to Dumoulin. The Brit would have most likely claimed the pink jersey.

Vuelta a España

The model suggested that Tom Dumoulin would have beaten Chris Froome in stages 2, 4, 12, 18, 19 and 21. In spite of a surge by the Dutchman towards the end of the race, the red jersey would have remained with Froome.


Based on a Collaborative Filtering approach, the results of 2017 suggest that Chris Froome would have beaten Tom Dumoulin in any of the Grand Tours.

Deep Learning – Faking It

Screen Shot 2017-08-20 at 15.01.01
Thumbnails of real bikes (Bianchi, Giant, Cube…)
Screen Shot 2017-08-20 at 15.01.15
Fake thumbnails generated randomly by Wasserstein Generative Adversarial Network

My last blog showed the results of using a deep convolutional neural network to apply different artistic styles to a photograph of cyclist.  This article looks at the trendy topic of Generative Adversarial Networks (GANs). Specifically, I investigate the application of a Wasserstein GAN to generate thumbnail images of bicycles.

In the field of machine learning, a generative model is a model designed to produce examples from a particular target distribution. In statistics, the output might be samples from a Gaussian distribution, but we can extend the idea to create a model that produces examples of sonnets in the style of Shakespeare or pictures of cats… or bicycles.

The adversarial framework introduces an attractive idea from game theory: to create a competitive form of learning. While a generator learns from a corpus of real examples how to create realistic “fakes”, a discriminator (or critic) learns to distinguish been fakes and authentic examples. In fact, the generator is given the objective of trying to fool the discriminator. As the discriminator improves, the generator is driven to enhance the authenticity of its output. This creates a virtuous cycle.

When originally proposed in 2014, Generative Adversarial Networks stimulated much interest, but it proved hard to make them work reliably in practice. One problem was “mode collapse”, where the generator becomes stuck, producing the same output all the time. However, this changed with the publication of a recent paper, explaining how earlier problems could be overcome by using a so-called Wasserstein loss function.

As an experiment, I downloaded a batch of images of bicycles from the Internet. After manually removing pictures with riders and close-ups of components, there were about 1,200 side views of road bikes (mostly with handlebars to the right, so you can see the chainset). After a few experiments, I reduced the dataset to the 862 images, by automatically selecting bikes against a white background.

Screen Shot 2017-08-20 at 14.45.29
Sample of real bike images

As a participant of part 2 of the excellent fast.ai deep learning course, I made use of WGAN code that runs using Pytorch. I loaded the bike images at thumbnail size of 64×64 (training with larger images exceeded the memory constraints of the p2.large GPU I’m running on AWS). It was initially disappointing to experience the mode collapse problem, especially because the authors of the WGAN paper claimed never to have encountered it. However, speeding up the learning rate of the generator seemed to solve the problem.

Although each fake was created from a completely random starting point, the generator learned to produce images against a white background, with two circles joined by lines. After a couple of hundred iterations the WGAN began to generate some recognisably bicycle-like images. Notice the huge variety. Some of the best ones are shown at the top of this post.

Screen Shot 2017-08-20 at 14.41.19
Sample of images generated by WGAN

I tried to improve the WGAN’s images, using another deep learning tool: super resolution. This amazing technique is used to solve the seemingly impossible task of converting images from low resolution to high resolution. It is achieved by taking downgraded versions of a large dataset of high resolution images, then training a neural network to reproduce a high-res version from the corresponding low-res input. A super resolution network is able to learn about certain properties of the world, for example, it converts jagged curves into smooth ones – a feature I’d hoped might be useful for making wheels look rounder.

Example of a super resolution network on real photographs

Unfortunately, my super resolution experiments did not lead to the improvement I’d hoped for. Two possible explanations are that a) the fake images were not low-res photos and b) the network had been trained on many types of images other than bicycles with white backgrounds.

Example of super resolution network on a fake bicycle image

In the end I was pretty happy with the best of the 64×64 images shown above. They are at least as good as something I could draw by hand. This is an impressive example of unsupervised learning. The trained network is able to use some learned notion of what a bicycle looks like in order to produce new images that possess similar properties. With more time and training, I’m sure the WGAN could be improved, perhaps to the point where the images might provide creative inspiration for new bike designs.


Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., … Bengio, Y. (2014). Generative Adversarial Networks. 

Arjovsky, M., Chintala, S., & Bottou, L. (2017). Wasserstein GAN. 

Johnson, J., Alahi, A., & Fei-Fei, L. (2016). Perceptual Losses for Real-Time Style Transfer and Super-Resolution. 


Deep Learning – Cycling Art

I’ve always be fascinated by the field of artificial intelligence, but it is only recently that significant and rapid advances have been made, particularly in the area of deep learning, where artificial neural networks are able to learn complex relationships. Back in the early 1990s, I experimented with forecasting share prices using neural networks. Performance was not much better than the linear models we were using at the time, so we never managed money this way, though I did publish a paper on the topic.

I am currently following an amazing course offered by fast.ai that explains how to programme and implement state of the art techniques in deep learning. Image recognition is one of the most interesting applications. Convolutional neural networks are able to recognise the content and style of images. It is possible to explore what the network has “learnt” by examining the content of the intermediate layers, between the input and the output.

Over the last week I have been playing around with some Python code, provided for the course, that uses a package called keras to build and run networks on a GPU using Google’s TensorFlow infrastructure. Starting with a modified version of the publicly available network called VGG16, which has been trained to recognise images, the idea is to combine the content a photograph with the style of an artist.

An image is presented to the network as an array of pixel values. These are passed through successive layers, where a series of transformations is performed. These allow the network to recognise increasingly complex features of the original image. The content of the image is captured by refining an initially random set of pixels, until it generates similar higher level features.

The style of an artist is represented in a slightly different way. This time an initially random set of pixels is modified until it matches the overall mixture of colours and textures, in the absence of positional information.

Finally, a new image is created, again initially from random, but this time matching both the content of the photograph and the style of the artist. The whole process takes about half an hour on my MacBook Pro, though I also have access to a high-spec GPU on Amazon Web Services to run things faster.

Here are some examples of a cyclist in the styles of Cézanne, Braque, Monet and Dali. The Cézanne image worked pretty well. I scaled up the content versus style for Braque. The Monet picture confuses the sky and trees. And the Dali result is just weird.



Trained to Forecast – Risk Magazine, January 1993

Deep Learning for Coders

A Neural Algorithm of Artistic Style, Leon A. Gatys, Alexander S. Ecker, Matthias Bethge