Blog

Despite impressive advances in machine learning, recent claims that AI can replace physical laws are deeply flawed. Machine learning models often just replicate existing knowledge from data generated by known physics, lacking true understanding or causal reasoning. While AI is a powerful tool for accelerating research and automating tasks, genuine scientific discovery and the formulation of physical laws still require human insight and expertise.

Learn how to implement gradient accumulation in PyTorch to train larger models on limited GPU resources, improving model performance and training stability.

A playful rebranding of machine learning jargon with a physics twist—plus a shoutout to Geoff Hinton’s historic achievements!

Our temporal predictive coding model provides a biologically plausible explanation for how the brain processes and predicts dynamic sensory inputs, bridging the gap between normative theories and neural mechanisms.

153 years ago, Charles Babbage—the father of the computer—passed away. His design for the Analytical Engine laid the foundations for modern computing, introducing core ideas like memory, programmability, and automatic calculation.

This blog offers a comprehensive technical and philosophical exploration of Kalman filtering, focusing on its implementation and reinterpretation within the PyTorch ecosystem. I will explore the mathematical foundations, practical coding strategies, differentiable filtering, and the deeper questions about knowledge, uncertainty, and learning that arise when classical estimation theory meets modern neural computation.

Normalizing Flows might not be the flashiest generative models, but their ability to provide exact, normalized densities makes them invaluable for tasks like variational inference. The new ‘Tensorizing Flows’ approach upgrades NFs with a powerful tensor-train base, making them far better at handling complex, high-dimensional, and multimodal distributions—especially in scientific and physics applications.

This report proposes a straightforward extension of the predictive coding model to handle dynamic (time-varying) stimuli, using a neural network architecture that remains biologically plausible and simple. The model can infer hidden variables and learn their dynamics from sensory input, as demonstrated on a toy example, and suggests directions for further development and application.

In this blog, I am going to provide a summary of a book chapter that I recently read, ‘Allostasis, interoception, and the free energy principle: Feeling our way forward’ by Andrew W Corcoran & Jakob Hohwy. I am also going to give you my own perspective on the topic and form some questions that could be interesting for future research.