No, we also provide a deep learning appliance that can be installed in customer data centers. Inquire at email@example.com for more information.What model architectures are currently supported?
Deep neural networks with convolutional, ReLu, maxout, or sigmoid units. Recurrent Neural Networks. T-SNE (a clustering algorithm)What data formats are currently supported?
CSV data, time series data, and image data. Generally available video and text coming soon.What fancy deep learning features do you provide?
Model ensembling, dropout averaging, Nesterov momentum, parameter annealing, and probably some others that improve accuracy and save on implementation time.How do I set hyperparameters?
We provide functionality that uses machine learning to set the hyperparameters for you, based on models that have trained well on similar data in the past. We still let you set model parameters if you want, but we can also pick them for you.How much data can I use?
Theoretically, it's unlimited because data is read in batches instead of all at once. In practice, you're limited by model capacity (IE number of neurons). Deep learning scales naturally to very large amounts of data, and this is one of the attractive aspects of the technique. With deep learning, "too much data" becomes a high quality problem.Who is the target user? Data scientists or marketing managers?
Our stereotypical user is an engineer with basic knowledge of machine learning. If you've watched the first few videos from Andrew Ng's Coursera class and know what an API is, you know enough to use Ersatz to get work done.Do you use GPUs?
Yes, we use GPUs (Graphics Processing Units) to speed up the training time of our models by up to 30x.