This paper describes three machine learning contests held as part of the ICML 2013 workshop on Challenges in Representation Learning. The workshop aimed to explore the latest developments in representation learning, with a focus on testing the capabilities of current algorithms through these contests. The contests included the black box learning challenge, the facial expression recognition challenge, and the multimodal learning challenge. The paper summarizes the datasets used, the results of the competitions, and provides suggestions for future challenges. The black box learning challenge involved using obfuscated data to test algorithms' ability to benefit from extra unsupervised data. The dataset was an obfuscated subset of the Street View House Numbers dataset. Competitors used various methods, including sparse filtering, random forests, and support vector machines. David Thaler won with an accuracy of 70.22% using a combination of these methods. The challenge highlighted the effectiveness of semi-supervised learning techniques. The facial expression recognition challenge aimed to recognize emotions in facial images. The dataset contained 35,887 images of various emotions. Competitors used convolutional neural networks, achieving high accuracy. Ian Goodfellow found that human accuracy on this task was around 68%, while the best model achieved 65.5%. The results suggested that convolutional networks can outperform hand-designed features. The multimodal learning challenge involved combining images and text data. Competitors used the ESP game dataset for training. The challenge tested the ability of algorithms to discover a unified semantic representation. The contest ended in a three-way tie with 100% accuracy. The winners used different approaches, including deep learning models. The paper concludes that competitions offer a different and important viewpoint on machine learning algorithms compared to research papers. They provide a realistic evaluation of generalization error and can help refocus attention on effective algorithms. The results of the contests highlight the performance of SVM loss functions, sparse filtering, and entropy regularization. The paper provides advice for future contest organizers, emphasizing the importance of designing fair and challenging contests.This paper describes three machine learning contests held as part of the ICML 2013 workshop on Challenges in Representation Learning. The workshop aimed to explore the latest developments in representation learning, with a focus on testing the capabilities of current algorithms through these contests. The contests included the black box learning challenge, the facial expression recognition challenge, and the multimodal learning challenge. The paper summarizes the datasets used, the results of the competitions, and provides suggestions for future challenges. The black box learning challenge involved using obfuscated data to test algorithms' ability to benefit from extra unsupervised data. The dataset was an obfuscated subset of the Street View House Numbers dataset. Competitors used various methods, including sparse filtering, random forests, and support vector machines. David Thaler won with an accuracy of 70.22% using a combination of these methods. The challenge highlighted the effectiveness of semi-supervised learning techniques. The facial expression recognition challenge aimed to recognize emotions in facial images. The dataset contained 35,887 images of various emotions. Competitors used convolutional neural networks, achieving high accuracy. Ian Goodfellow found that human accuracy on this task was around 68%, while the best model achieved 65.5%. The results suggested that convolutional networks can outperform hand-designed features. The multimodal learning challenge involved combining images and text data. Competitors used the ESP game dataset for training. The challenge tested the ability of algorithms to discover a unified semantic representation. The contest ended in a three-way tie with 100% accuracy. The winners used different approaches, including deep learning models. The paper concludes that competitions offer a different and important viewpoint on machine learning algorithms compared to research papers. They provide a realistic evaluation of generalization error and can help refocus attention on effective algorithms. The results of the contests highlight the performance of SVM loss functions, sparse filtering, and entropy regularization. The paper provides advice for future contest organizers, emphasizing the importance of designing fair and challenging contests.