This is a handwritten digit recognition competition.
Data
<p>Each sample is a 24x24 binary image. We vectorize each image as a 1x576 vector and stack N samples to represent the dataset as a matrix X of size Nx576. Each label takes value y in the set {0,1,2,...,9}.</p>
Data
<p>Sometimes it will be convenient to represent each label as a one-hot-vector, with one element set to one and the rest zero. The dataset labels can then be represented as a matrix, Y \in {0,1}^{Nx10} such that the sum of each column adds up to 1.</p>
File format
<p>
mnist/train/data
mnist/train/labels 
mnist/val/data 
mnist/val/labels
mnist/kaggle/data
</p>
There are three datasets:
large_train: training data (7000 samples)
val: validation data (2000 samples)
kaggle: kaggle competition test data (1000 samples, no labels provided)
You should train on the training data and validate your models using the validation data. Once you are happy with your performance, you should predict the samples in kaggle and submit the results (see Evaluation).