Imagine trying to fit an elephant into a suitcase.
Tough, right?
That’s kind of what AI researchers are doing—making massive AI models smaller and more efficient without losing their smarts.
What’s the Big Deal?
AI is getting ridiculously good at tasks like recognising faces, analysing medical scans, and even writing essays (yes, I see the irony) and doing your homework.
But here’s the problem: these AI models are getting huge—too big for smaller devices like smartphones or smart cameras.
They guzzle electricity like a teenager on energy drinks.
So, the brains at Terra Quantum AG have cracked the code on making AI smaller, faster, and less power-hungry.
Their latest research combines two superpowers:
- Hyperparameter Optimisation (HPO) – Tweaking AI settings to squeeze out the best performance.
- Neural Network Compression – Squashing AI models down without making them dumber.
Hyperparameter Optimisation: Finding AI’s Sweet Spot
Think of training an AI model like baking a cake.
You have to get the right mix of ingredients—flour, sugar, eggs, oven temperature—to make it perfect, otherwise it’s going to taste like my mothers baked goods and have the consistency or a large chalky rock.
You need something that ensures all the correct ingredients go into the cake in the right ratios to make your cake edible and delicious.
Hyperparameter optimisation (HPO) does the same for AI by fine-tuning settings like learning rate, batch size, and the number of neurons.
The researchers introduced TetraOpt, a new optimisation algorithm that’s better than the usual suspects—Bayesian Optimisation, Particle Swarm Optimisation, and Genetic Algorithms.
Who uses TetraOpt?
Banks, Defence Companies, IOT vendors, anyone that needs to squeeze the most of of AI.
In plain English: it finds the best AI settings faster and more accurately.
Neural Network Compression: Slimming Down AI Without Starving It
Deep learning models like ResNet-18 and ResNet-152 are big.
Think of them as huge encyclopaedias filled with AI knowledge.
But do you really need the entire encyclopaedia for every little question?
By using advanced mathematical techniques called tensor decompositions (CP, SVD, Tucker), these researchers shrank ResNet-18 by 14.5× and ResNet-152 by 2.5× while keeping their accuracy intact. The result?
✔️ AI models that run faster without needing a supercomputer.
✔️ AI that fits onto smaller devices.
✔️ Less energy consumption = happier planet.
Why Does This Matter?
Here’s where things get exciting:
- Edge AI & IoT: Imagine security cameras that can recognise faces on the spot, without sending data to the cloud. Faster, safer, and more private.
- Industrial AI: Factories can use AI to predict equipment failures, reducing costly downtime.
- Sustainability: Smaller, more efficient AI means lower power consumption—good for your electricity bill and the environment.
A Step Towards AI That Works Anywhere
This research builds on Terra Quantum AG’s earlier work on compressing large language models (LLMs) like GPT-2.
They’ve shown that tensor decomposition can shrink AI models while keeping them sharp—making AI more practical, private, and green.
With this new breakthrough, AI is getting lighter, faster, and ready to run anywhere—from tiny IoT sensors to industrial robots.
The future of AI isn’t just about being smarter—it’s about being more efficient.
And that changes everything. 🚀
