Recurrent nets are a powerful set of artificial neural network algorithms especially useful for processing sequential data such as sound, time series (sensor) data or written natural language. A version of recurrent networks was used by DeepMind in their work playing video games with autonomous agents.
Recurrent nets differ from feedforward nets because they include a feedback loop, whereby output from step n-1 is fed back to the net to affect the outcome of step n, and so forth for each subsequent step. For example, if a net is exposed to a word letter by letter, and it is asked to guess each following letter, the first letter of a word will help determine what a recurrent net thinks the second letter will be, etc.
This differs from a feedforward network, which learns to classify each handwritten numeral of MNIST independently according to the pixels it is exposed to from a single example, without referring to the preceding example to adjust its predictions. Feedforward networks accept one input at a time, and produce one output. Recurrent nets don’t face the same one-to-one constraint.
While some forms of data, like images, do not seem to be sequential, they can be understood as sequences when fed into a recurrent net. Consider an image of a handwritten word. Just as recurrent nets process handwriting, converting each cursive image into a letter, and using the beginning of a word to guess how that word will end, so nets can treat part of any image like letters in a sequence. A neural net roving over a large picture may learn from each region what the neighboring regions are more likely to be.
Recurrent nets and feedforward nets both “remember” something about the world, in a loose sense, by modeling the data they are exposed to. But they remember in very different ways. After training, feedforward net produces a static model of the data it has been shown, and that model can then accept new examples and accurately classify or cluster them.
In contrast, recurrent nets produce dynamic models – i.e. models that change over time – in ways that yield accurate classifications dependent of the context of the examples they’re exposed to.
To be precise, recurrent models include the hidden state that determined the previous classification in a series. In each subsequent step, that hidden state is combined with the new step’s input data to produce a) a new hidden state and then b) a new classification. Each hidden state is recycled to produce its modified successor.
Human memories are also context aware, recycling an awareness of previous states to properly interpret new data. For example, let’s take two individuals. One is aware that she is near Jack’s house. The other is aware that she has entered an airplane. They will interpret the sounds “Hi Jack!” in two very different ways, precisely because they retain a hidden state affected by their short-term memories and preceding sensations.
Different short-term memories should be recalled at different times, in order to assign the right meaning to current input. Some of those memories will have been forged recently, and other memories will have been forged many time steps before they are needed. The recurrent net that effectively associates memories and input remote in time is called a Long Short-Term Memory (LSTM), as much as that sounds like an oxymoron.
Recurrent nets have predictive capacity. They grasp the structure of data dynamically over time, and they are used to predict the next element in a series. Those elements might be the next letters in a word, or the next words in a sentence (natural language generation); the next number in data from sensors, economic tables, stock price action, etc.
Sequential data also includes videos, and recurrent networks have been used for object and gesture tracking in videos in real-time.
Recurrent nets, like other neural nets, are useful for clustering and anomaly detection. That is, they recognize similarities and dissimilarities by grouping examples in vector space and measuring their distance from each other. Modeling normal behavior, and flagging abnormalities, is applicable to healthcare data generated by wearables; home data generated by smart objects such as thermostats; market data generated by the movement of stocks and indices; personal financial data generated by account transactions (which may be used to identify fraud and money laundering).
Recall that Deeplearning4j’s multinetwork configuration lets you create a layer in the API simply by naming it. In this case, you create an LSTM.
In Deeplearning4j, normal LSTMs expect a matrix in which the first row, x_i, is given, and all subsequent rows, x_s, are what the neural network attempts to predict. This is a generative model, and there are no labels. There is no limit to the number of rows the matrix can contain, but all rows must have the same length.
The Graves LSTM, of which an example is forthcoming, is meant to be usesd in a multilayer network.
For example, input data could be:
Recurrent neural networks “allow for both parallel and sequential computation, and in principle can compute anything a traditional computer can compute. Unlike traditional computers, however, RNN are similar to the human brain, which is a large feedback network of connected neurons that somehow can learn to translate a lifelong sensory input stream into a sequence of useful motor outputs. The brain is a remarkable role model as it can solve many problems current machines cannot yet solve.” - Juergen Schmidhuber
Much research in recurrent nets has been led by Juergen Schmidhuber and his students, notably Sepp Hochreiter, who identified the vanishing gradient problem confronted by very deep networks and later invented Long Short-Term Memory (LSTM) recurrent nets, as well as Alex Graves, now at DeepMind. Two other researchers of note are Felix Gers, who invented LSTM forget gates, and Justin Bayer, who came up with a way to automatically evolve various kinds of LSTM topologies in a problem-specific fashion.