The paper discusses three machine learning contests held as part of the ICML 2013 Workshop on Challenges in Representation Learning: the Black Box Learning Challenge, the Facial Expression Recognition Challenge, and the Multimodal Learning Challenge. Each contest aimed to test and advance representation learning algorithms through practical applications and real-world datasets. 1. **Black Box Learning Challenge**: This contest focused on semi-supervised learning, using an obfuscated dataset from the Street View House Numbers (SVHN) dataset. The goal was to train classifiers without access to the original data's structure. The winner, David Thaler, achieved 70.22% accuracy using a combination of sparse filtering, random forests, and support vector machines. The contest highlighted the importance of leveraging unlabeled data and the effectiveness of simple techniques like entropy regularization. 2. **Facial Expression Recognition Challenge**: This contest aimed to compare feature learning methods with hand-engineered features on a new dataset. The dataset, FER-2013, was created using Google image search queries and included images of faces expressing various emotions. The winner, Yichuan Tang, used an SVM loss function with a convolutional neural network, achieving 100% accuracy. The contest demonstrated that convolutional networks can outperform hand-designed features, though the difference was not extreme. 3. **Multimodal Learning Challenge**: This contest focused on developing algorithms that can learn from multiple input modalities, specifically images and text. The challenge used a small ESP game dataset and a manually labeled test set. The contest ended in a tie with 100% accuracy, highlighting the difficulty of the task. The organizers suggested labeling twice as many test images as needed to make the problem more challenging. The paper also provides advice for future contest organizers, emphasizing the importance of time management, rule design, difficulty level, and multiple contests to ensure high participation and interesting results. Competitions offer a unique perspective on machine learning algorithms, allowing practitioners to explore different methods and evaluate their performance in realistic settings.The paper discusses three machine learning contests held as part of the ICML 2013 Workshop on Challenges in Representation Learning: the Black Box Learning Challenge, the Facial Expression Recognition Challenge, and the Multimodal Learning Challenge. Each contest aimed to test and advance representation learning algorithms through practical applications and real-world datasets. 1. **Black Box Learning Challenge**: This contest focused on semi-supervised learning, using an obfuscated dataset from the Street View House Numbers (SVHN) dataset. The goal was to train classifiers without access to the original data's structure. The winner, David Thaler, achieved 70.22% accuracy using a combination of sparse filtering, random forests, and support vector machines. The contest highlighted the importance of leveraging unlabeled data and the effectiveness of simple techniques like entropy regularization. 2. **Facial Expression Recognition Challenge**: This contest aimed to compare feature learning methods with hand-engineered features on a new dataset. The dataset, FER-2013, was created using Google image search queries and included images of faces expressing various emotions. The winner, Yichuan Tang, used an SVM loss function with a convolutional neural network, achieving 100% accuracy. The contest demonstrated that convolutional networks can outperform hand-designed features, though the difference was not extreme. 3. **Multimodal Learning Challenge**: This contest focused on developing algorithms that can learn from multiple input modalities, specifically images and text. The challenge used a small ESP game dataset and a manually labeled test set. The contest ended in a tie with 100% accuracy, highlighting the difficulty of the task. The organizers suggested labeling twice as many test images as needed to make the problem more challenging. The paper also provides advice for future contest organizers, emphasizing the importance of time management, rule design, difficulty level, and multiple contests to ensure high participation and interesting results. Competitions offer a unique perspective on machine learning algorithms, allowing practitioners to explore different methods and evaluate their performance in realistic settings.