Category: Model Optimisation Techniques

Hypernetwork Architectures

Hypernetwork architectures are neural networks designed to generate the weights or parameters for another neural network. Instead of directly learning the parameters of a model, a hypernetwork learns how to produce those parameters based on certain inputs or contexts. This approach can make models more flexible and adaptable to new tasks or data without requiring…

Neural ODE Solvers

Neural ODE solvers are machine learning models that use the mathematics of differential equations to predict how things change over time. Instead of using traditional layers like in standard neural networks, they treat the system as a continuous process and learn how it evolves. This approach allows for flexible and efficient modelling of time-dependent data,…

Differentiable Programming

Differentiable programming is a method of writing computer programs so that their behaviour can be automatically adjusted using mathematical techniques. This is done by making the entire program differentiable, meaning its outputs can be smoothly changed in response to small changes in its inputs or parameters. This approach allows computers to learn or optimise tasks…

Neural Combinatorial Optimisation

Neural combinatorial optimisation is a method that uses neural networks to solve complex problems where the goal is to find the best combination or arrangement from many possibilities. These problems are often difficult for traditional computers because there are too many options to check one by one. By learning from examples, neural networks can quickly…

Trust Region Policy Optimisation

Trust Region Policy Optimisation, or TRPO, is a method used in reinforcement learning to help computers learn how to make decisions. It works by ensuring that each learning step does not move too far from the previous strategy, which keeps learning stable and prevents sudden mistakes. By carefully controlling how much the computer’s decision-making policy…

Q-Learning Variants

Q-Learning variants are different versions or improvements of the basic Q-Learning algorithm, which is a method used in reinforcement learning to help computers learn the best actions to take in a given situation. These variants are designed to address limitations of the original algorithm, such as slow learning speed or instability. By making changes to…

Reward Shaping

Reward shaping is a technique used in reinforcement learning where additional signals are given to an agent to guide its learning process. By providing extra rewards or feedback, the agent can learn desired behaviours more quickly and efficiently. This helps the agent avoid unproductive actions and focus on strategies that lead to the main goal.

Invertible Neural Networks

Invertible neural networks are a type of artificial neural network designed so that their operations can be reversed. This means that, given the output, you can uniquely determine the input that produced it. Unlike traditional neural networks, which often lose information as data passes through layers, invertible neural networks preserve all information, making them especially…