Real numbers first
Self-attention, Q/K/V and multi-head attention computed by hand on a three-word sentence, so every formula has a number you can check.
AI courses by industry scientists with Ph.D.s
Learn AI,
from the math up.
LatentQuest courses teach the models behind modern AI the way they actually work: real numbers you can check on paper, then clean PyTorch, then real projects on real data.
Want a discount? Ask for a coupon →
- 17sessions
- NLP · Vision · Searchhands-on projects
- PyTorchbuilt from scratch
Courses
Learn AI the way it actually works
Every course follows the same path: work through the math with real numbers, build it in PyTorch, then ship real projects.
The course
Master Transformers from Scratch: BERT, ViT & AI Search
From vectors and attention math to BERT, Vision Transformers and two-tower semantic search. Taught by Mohammad R. Khalili Shoja and Seyed Vahid Mirnezami. You won't just hear definitions. You'll see the numbers, build the layers, and ship code that runs.
Then build it in PyTorch
Tensors, autograd and nn.Module, then a BERT-style masked language model written layer by layer.
Then ship real projects
Text classification, named entity recognition, distillation, Vision Transformers and two-tower semantic search on real Kaggle data.
- 00 Vectors, Dot Products, and Matrices
- 01 Why Attention? From Ambiguity to Context
- 02 Self-Attention with Real Numbers free code
- 03 Making Attention Learnable: Query, Key, and Value free code
- 04 Multi-Head Attention free code
- 05 The Transformer Encoder Block free code
- 06 Positional Encoding free code
- 07 BERT: The Encoder in the Wild
Interview-ready intuition and working code, in one place.
A guide for every session, and runnable code for every project.
Try it
Self-attention, live, with real numbers
The running example from the course. Pick a word to see how much it attends to every other word, and the new context-aware vector it ends up with. Everything below is computed live in your browser.
Session 2: attention with no learned weights.
Focus weights for bank
Each row of the attention matrix sums to 1.
Full attention matrix
Rows = queries, columns = keys.
Vector before → after
The new vector is a weighted mix of every word.
Open source
Free code from the course
Readable, tested, and runnable in Colab or on Kaggle. Start here, then take the full course for the complete path.
self-attention-from-scratchSessions 2–4
Self-attention explained with real numbers, then built in NumPy and PyTorch. Tests check every hand-worked number.
transformer-encoder-pytorchSessions 5–6
The full encoder block from scratch, plus an experiment: 50% accuracy without positional encoding, 100% with it.
vision-transformer-cassavaSession 15
Fine-tune a Vision Transformer on Kaggle cassava leaf photos, from patches to a competition submission.
semantic-search-two-towerSession 16
Two-tower retrieval and cross-encoder re-ranking, built by hand from a plain transformer encoder.
Instructors
Taught by people who build these systems for a living
We both earned our Ph.D.s at Iowa State University and now build computer vision, recommendation and search systems in industry. LatentQuest is how we teach what we wish we'd been taught: understand every layer, then build it, then ship it.
Stop treating transformers as a black box.
Go from vectors and dot products to BERT, Vision Transformers and semantic search, with code you understand line by line.