LEARN HOW TO FORGE 2026 LEVEL NEURAL NETWORK AND BELONG TO THE...
HYDRANET TITANS!
Rebuilt from scratch for AV 2.0. Every module rebuilt. Every concept rethought. This is not an update... this is a brand new cutting-edge course!
- Totally understand modern Deep Learning via Hydranets (not the 2019 Hydranets, but 2026 HydraNets (what I call HydraNets 2.0, which build for End-To-End Learning)
- Build practical skills via 3 projects on HydraNets, including 2 where you'll have complete freedom to build your own algorithms COLLECTIVELY (we enable a community hub where you'll be able to publish your work and get access to other people's work)
- Get access to a brand new course teaching brand new techniques like "Dead HydraNets", you'll rebuild the new Autoware 2.0 architecture, build your own lane detection heads, and more...
Dear friend,
If one of your objectives is to work on Deep Learning for AV, and rather than simply running black boxes, or AI agents, or models that you don't understand... your goal is to manipulate and forge them; then this page will show you how.
Here's the story:
Back in 2021, I discovered HydraNets after watching a Tesla conference. It was a term I caught in one of their 500+ slide deck, describing their architecture as a multi-task learning network. The idea? Rather than one network producing one output, the network could run MULTIPLE tasks at once. For example, depth estimation, segmentation, and object detection, all at the same time!
I often see myself as a "late adopter", but this time, I immediately embraced the concept, and went full speed on learning and teaching HydraNets... To make the concept as popular as I could, I wrote half a dozen articles about it, I built 2 courses on HydraNets, had one featured and promoted by giants of Computer Vision like PyImageSearch, I had researchers teach HydraNets into my course, I held a 1,000$ challenge, all because I bet big on the idea that Multi-Task Learning would be THE way to build architectures from now on.
And I was right.
After a few years, the term HydraNet is everywhere online.
And I may have contributed to that. When I look online, I see my diagrams have been republished by hundreds of blog posts all over the internet, in English, French, Chinese, Spanish, and many other languages...
I see companies like Tesla, Nvidia, XPeng, mobileye, Waymo, Neolix, and hundred others using HydraNets.
I see students naming themselves "HydraNet Engineer" on LinkedIn. I almost see a "movement".
For example:
In 2025, Autoware, the foundation powering 1,000+ self-driving cars via their open-source software announced during 2 conferences transitioning to End-To-End Learning and AV 2.0.
A move that many companies are currently doing.
In their first talk, they publicly announced using a "HydraNet". In the second, they showed visuals from my blog!
My blog's material is some of the most quoted, translated and reused in the entire industry - in particular the material on HydraNets & AV 2.0; which has been reused from China to Autoware's End-To-End conferences (AutoSeg 34:56, AutowareCon 1:13:09)
All over the globe, Deep Learning Engineers were learning and transitioning to HydraNets architectures.
The reason?
What I found interviewing End-To-End Powered Companies
In the last two years, I interviewed 50+ companies in the AV world. I particularly focused on those applying novel End-To-End learning methods, and I noticed two things:
In the last two years, I interviewed 50+ companies in the AV world. I particularly focused on those applying novel End-To-End learning methods, and I noticed two things:
HydraNets are the ESSENCE of AV 2.0
Besides a few exceptions, nearly ALL companies using End-To-End in production do it via HydraNets. But not the same HydraNets my v1 course was teaching, it was a different HydraNets — one I will named HydraNets 2.0... designed for End-To-End Learning, representation learning, knowledge transfer, with novel architectures that went far beyond what I was teaching.
I was watching all these complex architectures, and realized it was ME who had to go back to school and learn from the industry... so I could share it with you.
My v1 course was not teaching the novel structures (Necks, Dead Heads, ...) and was NOT giving you capacity to build your own head enough.
This brings me to reason 2:
You Deploy or Forge, not just RUN
And this is important: Job offers today want engineers who can either deploy models (use quantization, CUDA, TensorRT, ...) or Forge Models; rarely just running open-source.
What do I mean by 'Forge'?
I don't mean research, I mean adapting. Companies have specific needs, where sometimes, there is no "one size fits all". Sometimes, they have only partial labels to train their networks. Sometimes, they have a very specific environment, like off-road — where it's hard to use public data. Sometimes they need a specific encoder, or number of FPS, or specific tasks, such as lane detection, or road curb.
Knowing how to run a model is NOT the job, it's the bare minimum... the Deep Learning Labs, the R&D Centers, the AV companies don't want runners, they want forgers.
So if you want to work be at the cutting-edge of the Industry, I would recommend learning AV 2.0, learning HydraNets, and learning how to forge these networks.
Meet HydraNets v2... also known as...
HYDRANET TITANS!
In this new course, you will learn how to forge Deep Learning models for Multi-Task Learning.
I want you to know how to work with the latest multi-task learning approaches, but I also want you to understand them enough to OWN the models, and ultimately build your own versions... because THAT is the skill they are looking for.
You will build 3 HydraNet projects that you will be able to show to recruiters, 2 of them being "collaborative" (a new concept we are bringing to Think Autonomous), and you will learn Autoware's model, that are and will power hundreds of self-driving cars in the world.
This means if a company is running with Autoware, or any other End-To-End architecture that looks like a HydraNet (and the majority do), you will be the #1 best positioned to help them.
So let's see what you'll learn in the course:
WHAT YOU'LL LEARN
- Why nearly all End-To-End architectures are HydraNets, and how to add near unlimited heads to a network
- The 4 primitives of Deep Learning tasks — how to reverse-engineer any architecture in seconds
- The dirty little secret about HydraNet training nobody talks about (why your model will "bleed" entire classes after training, and the exact fixes to stop it)
- Why naively adding task losses destroys your network — and how to balance them so every task actually learns
- A surprising discovery (from Google) explains why the most complex MTL optimization methods are essentially useless (you'll see it in fast-moving fields like AI too, where people promote 200 complex agents while in reality, one with enough context is enough - you'll learn this in module 1)
- Why novel DL architectures are all multi-modal — and the 3 levels of HydraNet mastery (most engineers who know multi-task are still at Level 1)
- Why a model with 95% accuracy can still be completely wrong, and the visualization technique that exposes it (many models overfit or simply associate patterns to outputs — like grass with cows; we'll see how to REALLY learn)
- How to build Autoware's New HydraNet 2.0 from scratch
Pause:
Something important here:
When I built the v1 course, it was based on a researcher's work who built an Encoder-Decoder architecture on Depth Estimation & Segmentation.
The course was great, but I realized that it was very difficult to "own" the model.
By this I mean, you could run it, but not necessarily modify it with any head you want, and so on...
In this v2, you will build a 2026 level architecture that has everything: Neck, Context, Heads, Backbone, and more... Knowing about this will give you complete freedom to engineer a network based on problems. You will be tasked to build advanced networks, and work alongside a community of learners who also work on building HydraNets.
BUT we will NOT build Autoware's exact models. You will build your own version, on your own task, for example, drivable area segmentation & depth. Here is an example of what you'll learn at the beginning of the course:
Project: You will build a modern Multi-Task Model for Drivable Area & Segmentation
Notice the model here isn't producing perfect output. It's the POINT. It's been trained on 10 epochs only — but built from scratch. In this course, you will NOT run models, you will FORGE models.
One of your task will be to even create your own architecture, so you can tell companies you not only know the novel approaches, you can also take control of them.
Concept: You will study Autoware's new model End-To-End
There is more, so let me continue:
- HydraNets 2.0: The 4 components (Backbone, Head, Context, Neck) to master to be at the level of E2E DL Research Labs
- How to build temporal lane detection heads (when you think about it, some tasks like lane detection should NOT be frame-per-frame, we'll see how to engineer a professional pipeline solving it)
- What happens when you disassemble an already built HydraNet (in this part, you will actually implement the changes yourself, and understand how to forge your own custom network)
- The 6 ways to build lane detection algorithms with Deep Learning (and which one do major companies like Tesla, Waymo, Autoware, XPeng and NVIDIA chose — WARNING: Because of a specific element, some have abandoned lane detection entirely, explained in the lesson)
- How to build Foundation Segmentation Model that detects objects it was never trained on using pure CNNs (and how to stop doing 2017 deep learning)
- The #1 skill that separates engineers who can modify a production architecture from those who can only run one (this is a collective project where you will not only modify your architecture, but also report it to a "hub" and benefit from everybody else's own experimentations)
- The "dead HydraNet" technique companies like Autoware use (and you will do it from scratch) to assemble E2E models quickly
- 3 Ways to Build Object Detection Heads
- And many more...
These last bullets again tell the whole story:
You will build your own network, and you will be task to "invent" your own architecture.
An example?
Here is one you can build that does Lane Line Detection, on top of the two tasks from before (notice how lane detection IMPROVED depth & segmentation? It's a transfer task I will teach too):
Concept: HydraNets 2.0 are NOT about OUTPUT, they're about REPRESENTATION LEARNING
Project: You will learn to "Extend" HydraNets to multiple heads, such as (here) Lane Detection
This course teaches the new way to build Deep Neural Networks.
It's NOT about Transformers (we won't even see them inside, to keep focus on the topic of multi-task), so if you "just" know CNNs, you'll still be good...
And it's NOT about VLMs, LLMs, Action-Models, or any of these TOOLS.
So it's not about the HOW... it's not even about the WHAT...
It's about the "WHO".
Because my goal here is NOT to teach you one more trick that may or may not survive the next summer. My goal is to teach you a way of being a Deep Learning Engineer.
I want you to have skills, and you will, but I also want you to be a Leader in Deep Learning.
ON TOP OF THE STUDENT WHO SHARED HIS NEW HYDRANET PROJECT AT THE TOP, HERE ARE MORE...
STUDENTS OF HYDRANETS V1
"This is a MAGNIFICENT practical course"
Before joining, my biggest obstacle is the time required to complete this course. But I enrolled, and I have discovered a network architecture for multitasking that I did not know how to implement. I loved the quality of the material and the practical content, and I liked to learn about possible practical application fields of multitasking.
This is a magnificent practical course for discovering real areas for the application of multitasking architectures. The didactic quality of the course and its material were great.
Xose Ramon Fernandez Vidal
This course exists nowhere else
Let me go straight to the point:
This course exists nowhere else.
Coursera and Udacity are NOT going to teach you AV 2.0.
A university's curriculum? Puah!
Most of thse were designed in 2019 and barely got a new slide each year.
HydraNets v2 is a cutting-edge course that contains over 5 years of experience in multi-task learning, features the work of many engineers, is one of my most research oriented, and has been taught to many companies offline, via seminars, live courses, and more... This means I not only have built the course, but I also REFINED it over years.
It is NOT perfect. As I said, it deliberately obstruct Transformers, Self-Supervised Learning, VLMs, or other tricks, because I teach these ideas separately.
Because we are just launching this course, we are making it the lightest price it will ever be — and on top of this, we allow you to pay in 2 installments.
If you want to work on Deep Learning x AV 2.0, this is where it happens.
HYDRANET TITANS
IMMORTAL MULTI-TASK LEARNING TECHNIQUES FOR AV 2.0 FORGERS
- HYDRANET TITANS
[299€ value]
- LIFETIME ACCESS & UPDATES
- INSTANT ACCESS TO MODULES
Enroll For: 299€
* Or 2 payments of 159€
Frequently Asked Questions
HydraNets v2 is longer than v1, but we managed to keep it a course you can go through in a few hours. You can count around 7-10 hours to complete.
We have many companies purchasing our Deep Learning courses. If you work in a company that could benefit from it, it is worth letting them know this product exists and can help (not just you but also) them.
This is a self-study online course, which contains videos, articles, drawings, paper analysis, code, projects, and more...
This course is an intermediate/advanced course, which means several prerequisites are needed, including:
- Coding in Python (intermediate level)
- Basic knowledge of Deep Learning — you should know what a CNN is and how training works
- PyTorch basics are a Plus, but not strictly required — the course will build from first principles
If you have been through the v1, or if you are an Edgeneer Land Member, you can qualify for a discount. Check your inbox, or course, or latest Fragment for the coupon code.
A note that this course is ENTIRELY NEW. There is only 1 lesson that overlaps with the v1 (the Face Prediction example in Module 1), all the rest is completely new and cutting-edge!
You will do a lot of training during this course, but it should all be doable on Google Colab free GPUs. Obviously, you can also work on your own GPUs, and we made it friendly in this sense.
First, you get a lot of exercises that are not just 'do exercise' -> 'watch solution' - but more collaborative. I will ask you to make design choices, you will debate them with me, with other students as well. You will also be able to access a group where you can ask questions; and you will be able to benefit from the projects of other students!
This is a unique collaborative experience we built ONLY for this course.
© Copyright 2026 Think Autonomous™. All Rights Reserved.