Recent content in Home Page on Daniel Bethell
Neural networks learn complex internal representations of their inputs, but these representations are usually difficult to inspect directly. In this post, I constrain an image classifier to a two-dimensional embedding space and interactively visualise how different training objectives organise that space over time.
Quantifying uncertainty to reject out-of-distribution/adversarially attacked inputs in crucial for deploying deep learning models in the real-world. Our proposed method, C-EDL, boosts robustness in Evidential Deep Learning by detecting conflict from input transformations, improving OOD and adversarial detection without retraining, while keeping high accuracy and low overhead.
Robots cannot navigate safely if their perception misses hazards. COPPOL solves this by conformalising segmentation outputs to give distribution-free guarantees on hazard coverage, then using these calibrated hazard maps to guide reinforcement learning. The result is safer navigation under imperfect perception, with higher hazard coverage, fewer unsafe violations, greater clearance, and strong…
Peer review is supposed to reward merit and novel research, but the data suggests the playing field at the Top AI conferences isn’t perfectly level. Using public review metadata from major AI conferences (e.g., ICLR), I compare acceptance rates across research areas, and check whether differences persist even at similar review scores. What do we find; some fields have a small bias towards…
Safe reinforcement learning is essential for safety-critical domains, but current methods are often rely on prior knowledge. We have developed ADVICE, a new post-shielding method that learns safety-relevant structure using a contrastive autoencoder, reducing safety violations while maintaining competitive reward, all without using any prior knowledge.
This tutorial dives into conformal prediction, covering basics, advanced methods with Python code, and resources to help you apply CP in your next project!
Kolmogorov-Arnold Networks have been promised to be the next big thing in deep learning. But what are they? How do they work? Do they live up to the claims? In this blog post, I will be trying to answer all of the following alongside easy to follow code and results.
Conformal Prediction is an amazingly powerful tool in safety-critical applications, but current state-of-the-art conformal prediction algorithms tend to overestimate their prediction sets/intervals. We have developed MC-CP, a new conformal prediction algorithm that reduces average set/interval size without any reduction in prediction performance.