Neural Networks & Architecture — Adex Krost
Where structure meets understanding
A transnational platform that started in 2023 with one idea — that geography should not determine who gets to study machine learning and deep architecture seriously.
What we actually do here
Adex Krost builds structured online lectures on neural network architecture — not survey courses, not introductory overviews. Each series goes deep into one area: how attention mechanisms distribute weight, why residual connections changed what depth can mean in practice, where transformer architecture diverges from earlier recurrent designs.
The student base spans over 40 countries. Some are postgraduate researchers fitting sessions around lab hours. Others are working engineers who already ship models and want to understand what they are shipping. The material is written to hold up for both.
Lectures are delivered sequentially — each session builds directly on the previous. That structure is deliberate. It produces the kind of understanding you can actually use when something breaks or behaves unexpectedly, rather than surface-level familiarity that evaporates under pressure.
Every concept is grounded in specific architecture choices — why GPT uses unidirectional masking, how the encoder-decoder bottleneck shapes translation quality, what the skip connections in UNet are compensating for. Abstraction without that kind of anchoring tends not to stick.
Dr. Oksana Bilyk
Lead Curriculum Architect
Countries with active students enrolled in current series
Sequential lecture series available, each independently complete
Completion rate across multi-session architecture series
How the material is structured
Three guiding principles shape every series — not as abstract ideals but as concrete decisions that affect what gets covered and in what order.
Sequential delivery
Sessions are ordered so that each one depends on the previous. Skipping ahead works against comprehension — the sequencing is load-bearing, not decorative.
Architecture specificity
Every claim connects to a named model, a paper, or a measurable behavior. Vague generalisations about how networks "learn" are replaced with precise descriptions of what happens in specific layers.
Location independence
All materials are asynchronous by design. A student in Nairobi and one in Warsaw access identical content at whatever pace their schedule allows — no cohort gates, no time-zone constraints.
Architecture diagrams
Training analysis
Attention mechanisms
- Convolutional architectures — from LeNet geometry to depthwise separable filters in MobileNet
- Recurrent structures — LSTM gating logic, vanishing gradients, and where they still outperform transformers
- Self-attention and positional encoding — how sequences get represented without recurrence
- Encoder-decoder patterns — bottleneck representations, skip connections, U-shaped designs
- Scaling laws — what changes when you increase parameters versus data versus compute
"The lecture on residual connections was the first time I actually understood why deep networks stopped being difficult to train — not that they stopped being difficult, but why the specific problem of gradient degradation had a structural solution."
— Yaroslav Fedoriv, graduate researcher
Deep architecture overview