I’m Silas Bempong, a Machine Learning Engineer from Ghana with a knack for building AI that tackles real-world challenges. My expertise spans deep learning, ...
A small-model replication sketch inspired by Anthropic's emotion-concept work. Anthropic's 2026 paper, Emotion Concepts and their Function in a Large Language Model , reports that Claude Sonnet 4.5 contains internal activation directions corresponding to emotion concepts such as calm , afraid , and desperate . The key claim is not that the model feels anything, but that these directions are…
I built Stormlog after hitting the same problem too many times: memory debugging in training jobs kept turning into a fragmented workflow of ad hoc counters, partial profilers, logs that disappeared with the run, and artifacts that were hard to revisit later. I wanted one toolkit that could track memory live, save telemetry, analyze it after the fact, bundle diagnostics, export visualizations, and…
The motivating question If you’ve ever tried to copy an object from one photo into another, you’ve seen the problem immediately: seams . Even with careful masking, edge feathering, and opacity tweaks, the composite still tends to look “pasted.” Our eyes are incredibly sensitive to discontinuities in edges and lighting , not just raw color differences. So here’s the driving question: Why does…
TL;DR: On a small semantic segmentation task (Oxford-IIIT Pet) trained on a MacBook, ImageNet pretraining clearly helped more than heavier data augmentation. Pretraining gave me ~+3.8 mIoU points on average with essentially the same training time. Heavy augmentation was almost a wash. Why I Ran This Experiment Semantic segmentation models are usually trained on big GPUs with big datasets. But most…
Python libraries like PyTorch, NumPy, and JAX are indispensable for handling multidimensional arrays (tensors). However, these libraries can sometimes lead to subtle, silent bugs, particularly with array broadcasting, that can go unnoticed until they cause significant issues. Enter JaxTyping , a library for type annotations and runtime checking of array shapes and data types, which enhances code…
In machine learning, numbers can rage like a storm, threatening to crash your model into oblivion. I’ve wrestled with this chaos, watching computations spiral out of control. But there’s a tool that tames these storms: LogSumExp. Like a conductor calming a chaotic orchestra, it brings order to numerical explosions, ensuring neural networks run smoothly. Let’s explore how it works, why it matters,…
For a long while, I've been trying to set up a personal space where I could document my learnings over the last 5 or so years, but I never got to it. Once upon a time, I used to actively blog and share things online, but since I started work, the light dimmed completely. And with that said, I have a lot of things on my mind that I can finally start channeling in my own space. If Darrow O'Lykos…