Four of the latest Learning Paradigms in Machine Learning, as of 2025.

(And how they are correlated).

1. Online Learning (aka Real-Time Learning or Streaming Learning), as used in Reinforcement Learning (RL), has been traditionally coined for learning from streaming data. It involves updating a model incrementally as new data arrives, without requiring the entire dataset to be stored or the model to be retrained from scratch.

• Thus, training in Online Learning is not on a fixed dataset all at once (batch learning). So, we can say that most of the learning by the RL agent happens online as the agent actively interacts with its time-dynamic environment.

Note that Online Learning is important in online financial trading where new market history keeps coming to an automated trading agent. Imagine the picture on the left being a company announcement, and the picture on the right being an automated trading RL agent.

2. Residual Learning is learning what to add. Now, while residual learning is typically batch learning and it improves a deep network training and accuracy (since it involves learning the difference between the input and output of a neural network layer), online learning simply adapts to new data in real-time.

3. Fast Learning uses the versatile Extreme Learning Machines (ELMs) to employ a single-hidden-layer architecture so as to address the computational limitations inherent in deep neural network-based approaches. So, it diverges from deep neural networks, unlike residual learning.

• A baseline ELM is a single hidden layer feed-forward neural network. But there have been improved algorithms; like I-ELM, CI-ELM, EM-ELM, EIR-ELM, etc. Liexin Cheng, Xue Cheng, and Shuaiqiang Liu wrote a 2025 paper based on ELMs, titled “Fast Learning in Quantitative Finance with Extreme Learning Machine.”

4. With the Federated Learning (FL) architecture, instead of ML models being computed on large centralized machines, they are distributed over local devices (such as mobile devices) for computation. This hadn’t been practically possible in the past since computational abilities of mobile phones were very limited for running any ML model. Similarly to blockchain, both technologies seek solutions at the level of the marketplace, rather than within single organizations. The challenge with FL is that data encryption increases calculation volumes.

• Online Learning is conducive to FL. Online Learning’s ability to update models with new data points complements FL’s iterative aggregation of model updates from clients. Each client can perform online learning locally, adapting to new data, and contribute updates to the global model.

Sources:

Google Images.

Internet articles.