Let's revisit the OpenAI's task (https://blog.openai.com/requests-for-research-2/). As stated there, we need to generate a dataset of random 100,000 binary strings of length 50. In other words, our training set will be formed of 100,000 examples and the recurrent neural network will accept 50 time steps. The result of the last time step would be counted as the model prediction.
The task of determining the sum of a sequence can be viewed as a classification problem where the result can be any of the classes from 0 to 50. A standard practice in machine learning is to encode the data into an easily decodable numeric way. But why is that? Most machine learning algorithms cannot accept anything apart from numeric data, so we need ...
