Recurrent Neural Networks (RNN)
Recurrent Neural Networks (RNNs) are a class of artificial neural networks designed to recognise patterns in sequences of data, such as time series, spoken language, or transactional histories. Unlike traditional neural networks, RNNs have a “memory”, as they retain information from previous steps in the sequence and use it to influence the current output. This makes RNNs especially useful for tasks where the order and context of data points matter, such as transaction monitoring for example.
At their core, RNNs process input data step-by-step, looping the output of one time step back into the network for the next time step. This feedback loop enables the network to model temporal dynamics, which is critical for making predictions based on sequences.
Sequential neural networks
-
-
RNN
Examples
Imagine a European retail bank that processes millions of transactions per day across digital and physical channels. In order to stay compliant, they need to monitor all these transactions and look for signs of fraud. However, in order to identify suspicious transactions in real time, the bank has to take into account the customer’s entire history and look for sudden changes in behaviour.
So, using an RNN, the bank builds a model that learns from the sequence of a customer's past transactions, such as amounts, locations, times, and merchant types. Unlike a traditional model that would treat each transaction independently, the RNN considers the order and context of events.
For example, if a customer always spends in Switzerland and suddenly a transaction occurs in another country at an unusual time of day, the RNN will detect the anomaly based on that behavioural shift over time.
FAQ
What are the main drawbacks of RNNs?
While RNNs do have an internal memory, long-term dependencies can cause issues. They are also slower to train due to their sequential nature.
How is an RNN different from a regular neural network?
A standard neural network processes fixed-size input and treats each input independently. In contrast, an RNN handles sequential data and maintains an internal memory of past inputs.
Are RNNs still used today, or have they been replaced by newer models?
While newer models like Transformers have outperformed RNNs in many areas (especially in language tasks), RNNs and their variants are still used in narrower, real-time, or resource-constrained applications, particularly in finance, speech processing, and signal analysis.
