NTK & Feature Learning Regimes A Unified View
How neural networks transition from kernel-like behaviour to genuine feature learning, and why this boundary is the most important question in modern deep learning theory.
Mobile developer.
AI artist. Deep learning
researcher. Human.
I'm Yacine Herrouel a mobile developer and independent deep learning researcher from Earth. I build apps for iOS and Android, train my own neural networks, and create generative AI art. I believe code and creativity are the same muscle. Also quietly obsessed with quantum physics and the idea that 1 + 1 ≠ 2.
Apps. Models.
Generative art.
Scroll to explore
✦ Published works
Books I've written
✦ Deep Learning
MLP Is All You Need
A practical, no-fluff guide to deep learning foundations. Master the math, intuition, and implementation of Multi-Layer Perceptrons the building block behind every modern neural network.
✦ LLMs · SLMs · Philosophy
The Philosophy of LLMs & SLMs
A deep, philosophical dive into how large and small language models really work not just the engineering, but the ideas beneath them. Concepts are grounded through examples drawn from quantum physics, not as the subject, but as a lens for thinking about uncertainty, superposition, and emergence in AI.
✦ NLP · Philosophy
Natural Language Processing Guide
A philosophical and technical exploration of how machines understand language. Bridging linguistic theory, cognitive science, and modern NLP for those who want to understand not just the how, but the why.
✦ About me
On building
✦ What I do
Everything I build,
research, and obsess over
All projects →
- Cross-platform mobile with Flutter
- Dart clean, typed, fast
- Flutter Web & Desktop
- On-device ML with TensorFlow Lite
- REST & GraphQL API integration
- Play Store & App Store publishing
- Performance optimisation & profiling
Flutter.
Flutter.
Flutter.
- Neural network architecture design
- Training from scratch no wrappers, no shortcuts
- Multi-Layer Perceptrons & backpropagation
- Attention mechanisms & transformers
- Computer vision & image generation
- PyTorch & JAX experimentation
- Independent research & paper writing
Train.
Think.
Publish.
- Diffusion model fine-tuning & prompting
- GANs & variational autoencoders
- Custom style transfer pipelines
- Midjourney, Stable Diffusion, ComfyUI workflows
- Neural style & latent space exploration
- AI-assisted visual storytelling
Diffuse.
Generate.
Create.
- User research & journey mapping
- Wireframing & rapid prototyping
- Design systems & component libraries
- Figma from sketch to handoff
- Accessibility & inclusive design
- Motion design & micro-interactions
Design.
Prototype.
Ship.
- Architecture planning & system diagrams
- State management & data flow design
- API contract design & versioning
- Scalable patterns clean, layered, testable
- Algorithmic thinking & complexity analysis
- Documentation & technical writing
Architect.
Structure.
Scale.
- Intermittent fasting discipline as a daily practice
- Long-distance running & 10K training
- Amateur bodybuilding & progressive overload
- Physical rigour as a mental framework
- Recovery, rest, and knowing when to stop
- The overlap between sport and shipping software
Fast.
Run.
Lift.
✦ What drives me
Built with code,
curiosity, and iron
I don't separate what I do for work from what I do for myself. Whether I'm debugging a neural net at 2am, generating a piece of AI art, or finishing a 10K I'm always after the same thing: understanding something deeply and then pushing it further.
I'm an independent researcher because I believe the most interesting questions in deep learning don't have job titles. And I lift weights for the same reason I study quantum physics because hard things are worth doing.
Get in touchMobile apps shipped on iOS & Android✦
Independent deep learning research no institution required✦
AI-generated art series exploring latent space & human perception✦
Amateur bodybuilder. Runner. Recovering from rest days.✦
Obsessed with quantum physics and why 1 + 1 ≠ 2✦
From Earth building things that actually matter✦
Training neural networks from scratch every single layer✦
0
Mobile apps shipped across iOS & Android
🧠
Independent deep learning researcher no PhD required
5K+
Hours training, running, and lifting since day one
∞
Questions about quantum physics still unanswered
✦ Tools & stack
Technologies I build
and research with
✦ Currently working with
Manzilak
AI Systems & Mobile Dev
[email protected]
S****
AI Systems & Mobile Dev
[email protected]
Personal
Research · General inquiries & freelance
[email protected]
Swift
Kotlin
PyTorch
SwiftUI
JAX
Flutter
Core ML
TensorFlow
Python
React Native
NumPy
Xcode
Stable Diffusion
ComfyUI
Hugging Face
Android Studio
Git
Docker
Scikit-learn
Firebase
Pandas
Matplotlib
Jetpack Compose
ONNX
✦ Thoughts on paper
I write on Medium
01
↗
02
↗
The Only Phase of Scientific Research in AI
Most AI "research" is engineering with a hypothesis. Here's what it actually means to do science in a field where the ground truth keeps moving under your feet.
03
↗
Regularization, Gradients & Training Dynamics
A deep look at how regularization isn't just a trick to prevent overfitting it's a shaping force on the entire trajectory of learning, from first epoch to convergence.
04
↗
The Beauty of Imperfect AI
Perfection in a model isn't the goal calibrated uncertainty is. Why I've come to believe that a model that knows what it doesn't know is more valuable than one that pretends otherwise.
05
↗
Your Dataset Is Your AI Model Designer
The architecture gets all the credit, but the dataset makes all the decisions. A study of how data geometry encodes inductive biases that no hyperparameter sweep can override.
✦ Deep Learning Research
I work on the mathematical foundations of deep learning training dynamics, generalisation theory, feature formation, and the geometry of loss landscapes. I am also deeply interested in SLMs and LLMs, exploring how scale, architecture, and training shape emergent capabilities. The work is independent, rigorous, and conducted outside any institution.
I don't publish ideas before they're ready, and I don't expose work in progress on a public page. If you're a researcher, engineer, or someone genuinely curious about what I'm building reach out directly. I'm selective, but always open to the right conversation.
Reach me for research details
I respond to serious inquiries. Tell me what you're working on.
Neural Tangent Kernel
Feature Learning
Training Dynamics
Loss Landscape Geometry
Implicit Regularization
Overparameterization
Diffusion Models
Generalisation Theory
Dataset Design
On-Device ML
✦ Who I am
"
A remarkably ambitious developer whose technical knowledge outpaces others. Working closely with him on many projects, I have consistently been impressed by his precision and delivery. He was and still the person whom you can depend on no matter how tough or complicated the task is.
Abdellah Alghamedi · Owner at Manzilak
"