It’s Possible to Win ML Hackathon if You Ranked 2nd on Leaderboard

It’s Possible to Win ML Hackathon if You Ranked 2nd on LeaderboardRustem GaliullinBlockedUnblockFollowFollowingApr 24In machine learning, a high accuracy model is a key for success.

Nonetheless, models are build for end users.

This story will tell you how an end user perspective can add valuable points to succeeding in data science.

IntroA few weeks ago we took part in a machine learning hackathon hosted by Skoltech research institution.

It took place in our amazing hometown Kazan in Russia and lasted almost 2 days.

Definitely check out this nice portal about our home region below.

Visit Tatarstan – The Official Tourist Portal of the Republic of TatarstanWalk Like a Local.

5 Ways to Experience Kazan From the Inside More about the guidebook 100 Visit Tatarstan souvenirs to…visit-tatarstan.

comNote: spoilers ahead.

We ranked the second on the private leaderboard but we still won it, and I’m going to tell you how.

Hackathon OverviewThe hackathon had three tracks:Web-based labelling tool for doctorsMachine learning algorithm to classify types of epilepsyOpen track for any solutions within the AI in Medicine domainAs data scientists, we unsurprisingly picked the second track.

Problem DescriptionThe task was two-fold:Build a model that would correctly classify 8 types of epilepsy from brain EEG signals.

Build a web-app that inputs a file and spits out a prediction for an end user.

Table 1.

Types of Epilepsy SeizuresThere were 2012 observations collected from patients that had suffered an epilepsy seizure.

Each observation contained time-based samples and preprocessed using two popular methods.

More details on data preprocessing in this paper.

Figure 1.

Raw EEG Signals (tribute to WikiMedia Commons)After preprocessing, we received two files per each observation:3-dimensional array of shape [Number of Samples, Number of Channels, Number of Frequencies]2-dimensional array of shape [Number of Samples, Number of Frequencies]BenchmarkThe testing protocol was mean weighted F1-score on 5-fold cross-validation split on a file-level, i.


files are split into five folds, only then they must be parsed, because samples within one file are highly correlated.

If you’re wondering how to compute cross-validated F1-weighted score, here’s python code for that:import numpy as npfrom sklearn.

metrics import f1_score# assuming you have fold-wise true labels and predicted labels # as a list of tuplesval_scores = [f1_score(true_labels, predicted_labels, average=”weighted”) for true_labels,predicted_labels in fold_labels_and_predictions)]mean_f1 = np.

mean(val_scores)The authors of the paper took eigenvalues of matrices on sample-level and trained various classifiers, among which K-Nearest Neighbors beat other algorithms.

They were able to reach 88.

4% score using method 1 and 90.

7% using method 2.

LeaderboardAs for the leaderboard, our local 5-fold CV served as a proxy for the public score.

As for the private scores, we would find out them only after demoing our solutions.

However, there was a very important note from the host in the beginning.

Teams with highest scores would be ranked not only by private scores, but also by the quality of a web-app.

Neural NetWe decided to use a Convolutional Neural Network since many researchers proved it to be a good fit for a seizure detection tasks.

Also, we figured that EEG signals collection process is impacted by a topology of placing scanners on a human head.

Luckily, the first type of preprocessing provided suitable inputs for 2D convolutions.

ArchitecturesWe tried out different backbones for our convnet:pretrained Imagenet models (ResNet18, VGG11)Imagenet architecture with random weights (also ResNet18, VGG11)a simple CNN from scratchObviously, models pretrained on Imagenet didn’t perform well as inputs had nothing in common.

ResNet and VGG architectures with random weights didn’t outperform a simple convnet as well.

I tie this to the fact that spatial dimensions of our input are quite small (20 pixels in height and 24 pixels in width), whereas first convolutional layer of these models have a filter size of 7.

We end up with a bunch of lost signal and sometimes with negative output dimensions (as with AlexNet).

That being said, after long-lasting experimentation we settled with this tiny model:3 blocks of convolution, leaky ReLU, batch normalization and max pooling1 block of convolution, leaky ReLU and batch normalization1 fully-connected layerFigure 2.

Our CNN ArchitectureInputsSince we used a dataset of the method 1 preprocessing, our inputs were of shape [num_samples, 20, 24], whereas num_samples could range from 3 to over 100.

Thus, we tried different inputs:a random window of consecutive samples with window size being equal to 1, 3, 5 and 10mean value across all samples, i.


input shape is [1, 20, 24]adding standard deviation, maximum, minimum along with meanadding 10-th, 25-th, 50-th, 75-th and 90-th percentiles along with meanNo long talks, a random window didn’t work at all.

So, we took average across all samples and training got much more interesting, we reached almost 90% in our metric.

Adding other statistical measures didn’t seem to help, usually our score went down.

However, 25-th and 75-th percentiles gave us 1% boost, and we kept them making our input of shape [3, 20, 24].

Softmax layerI know we’re supposed to have a softmax layer in multi-class problems.

Well, we didn’t.

First, I forgot to add it and when measuring accuracy I was just finding an index with maximum value.

However, when I added softmax after discovering this negligence, model performance dropped.

It didn’t train as good as before and I removed it for good.

Training ResultsA few iterations and we figured out that Adam optimizer with learning rate between 0.

002 and 0.

005 gave us highest performance adding extra 1% in the cross-validation scores.

Each fold was trained for 40 epochs.

Final local performance was measured as an average over all 5 folds.

We hit 92.

39% with one-fold maximum at 93.


After all teams demoed their results, we were on the second place, 1% behind the leader.

Upon releasing private scores, our performance jumped to 94.


Sadly, we remained 1% behind.

Web AppNow, you should remember that there was one more criterion to winning.

Yes, a web application where a doctor can submit a file and get a prediction.

Table 2.

Sample Model OutputWe decided to split efforts to develop a nifty web application form.

Don’t hesitate to download a few example files from our drive to test this form.

It’s easy.

Just upload one or few files to get predictions for each file.

You can also download predictions as a CSV file.

Model InterpretabilityI hope you’ve spent a minute to check out our app.

Provided you have, you should have noticed an unusual chart like the one below.

Figure 3.

Grad-CAM Output InterpretationWe recalled that in medicine it’s important not only have a very high precision, but also a succinct explanation.

That’s why we spent time to bring Gradient-weighted Class Activation Mapping to give our users a tool to explain why our model makes any of its decisions.

Shortly, it uses a class prediction, backpropagated gradients and class activation maps to point out areas of interest on the input.

That’s how we get these two images above.

The left one is a heatmap of means across samples a.



input matrix, the right one is output of the Grad-CAM algorithm drawing a doctor’s attention to most influential input values.

Winning SolutionWell, that’s whole secret sauce of ours.

We blended high accuracy model with a simple yet complete web-app and flavored everything with visual model interpretation.

Even though, we were runner-ups in terms of weighted F1-score, the hackathon jury appreciated the fact that we considered how doctors could use this solution most efficiently.

Also, we utilized lessons learnt from a previous hackathon that it was quite important to have a unite team to work on one problem only.

We had a good mix of specialties and didn’t waste our limited time and energy on other tasks.

More about TeamI (Rustem Galiullin) — machine learning researcher — worked on a model performance.

Almaz Melnikov — machine learning engineer — filled our web-app with functionality and integrated our neural net.

Konstantin Yarushkin — front-end developer — developed a template for our web-app.

Our smiley faces after winningResourcesProject web-app https://deepl.


com/Our gitlab repo https://gitlab.


gitGrad-CAM implementation in pytorch https://github.

com/jacobgil/pytorch-grad-camGrad-CAM project http://gradcam.


org/About Skoltech https://www.


ru/en.. More details

Leave a Reply