Toy problem for RNNs in computer vision

Question

I am writing some neural network training code that I will later use to solve a specific problem. First, though, I want to validate that the code is working and is able to train an RNN (recurrent neural network) as expected.

I'm looking for a simple problem and dataset that I know a simple network can do very well on. Here are some guidelines/criteria:

• The problem should not be too difficult. I want to be able to train networks quickly so that I can fix bugs in the code quickly, without waiting several says for the training to complete. Thus, small networks should be able to solve the problem well.
• Ideally, reasonably small networks would be able to achieve near-perfect accuracy. This gives me a metric I know I should be able to aim for.
• The problem should not be trivial. Several resources online suggest generating simple number sequences, but I worry that my training code might break down when I switch to my real dataset.
• A computer vision based task would be preferable, since my real problem is one of action recognition. I hesitate to use datasets as complex as Sports1M and friends, though, since work on datasets like these is still the cutting edge--hardly a toy problem.

If I were implementing pure CNNs, I would simply use MNIST or CIFAR, but I'm unsure what the analog is for convolutional RNNs. What are some datasets that I might use?

Answer 1

This dataset comes up by search the UCI Machine Learning Repository for RNN (recurrent neural network)

Wall-Following Robot Navigation Data

The provided files comprise three different data sets. The first one contains the raw values of the measurements of all 24 ultrasound sensors and the corresponding class label (see Section 7). Sensor readings are sampled at a rate of 9 samples per second.

The second one contains four sensor readings named 'simplified distances' and the corresponding class label (see Section 7). These simplified distances are referred to as the 'front distance', 'left distance', 'right distance' and 'back distance'. They consist, respectively, of the minimum sensor readings among those within 60 degree arcs located at the front, left, right and back parts of the robot.

The third one contains only the front and left simplified distances and the corresponding class label.

And specific to RNN:

If a recurrent neural network, such as the Elman network, is used to learn the task, the resulting dynamical classifier is able to learn the task using less hidden neurons than the MLP network.