Machine Unlearning: teaching AI to forget on demand — with governance

The core dilemma of modern AI is that it doesn't store information like a traditional database, where you just find a row and hit delete. During training, data gets diluted and woven into millions of synaptic connections (the model's weights). Knowledge about any single individual ends up fused into the network's global intelligence.

How do you extract the sugar after it's already dissolved in the coffee?

The industry's default answer has been to throw out the coffee and brew a new pot from scratch. In other words: retrain the entire model without that customer's data. For production models, that means hours or days of compute, massive cloud infrastructure costs, the risk of degrading overall performance, and a response time that's completely incompatible with real-world agility.

Today's AI knows too much, but has terrible selective memory. That's exactly the governance problem we're tackling with Machine Unlearning — one of our core research fronts at Neuro.

What if there was a way to make AI forget just one specific data point, forcing the neural connections that hold that memory to collapse — without destroying the rest of its knowledge?

At Neuro, we implemented a technique based on controlled data poisoning via Label Flipping. Instead of hiding the target data from the model and retraining everything, we expose it to that same data, but with a deliberately contradictory label.

We run a quick fine-tuning cycle where we present the exact same features of the data that needs to be forgotten, but paired with a completely opposite or random category. It's the equivalent of taking a firmly established concept in someone's mind and, through intense contradictory repetition, forcing the brain to doubt that certainty.

In just a few iterations, the original pattern dissolves.

To automate and find the exact dosage of this "poisoning" (how many iterations, at what intensity), we use Keras Tuner with random search. We optimize based on a custom metric: the Unlearning Score (retained global accuracy minus the recognition rate of the data we need to forget). The higher the score, the cleaner the removal.

To validate the surgical efficiency of this architecture, we ran batteries of automated experiments on computer vision benchmarks (MNIST dataset), isolating and erasing entire data categories while preserving the rest of the ecosystem.

The results prove that programmed amnesia is viable and highly efficient:

96.2% Unlearning Score — achieved with just 5 epochs of focused fine-tuning.

~97% global intelligence retained — the model maintained excellent performance recognizing all non-target data categories, while recognition of the forgotten data dropped to near zero.

Under 15 seconds — the total processing time, eliminating 100% of the need for retraining from scratch.

Label Flipping is one of our entry-level tools, but the Machine Unlearning ecosystem attacks the mechanics of forgetting from several mathematical angles. We're already mapping and planning the implementation of even more sophisticated approaches for our pipeline, balancing speed, precision, and privacy.

Force Inversion (Gradient Ascent): reverse-engineers the learning process. Mathematically maximizes error only on the target data subset, forcing the network to actively move away from that knowledge.

Synaptic Surgery (Fisher Forgetting): uses statistical matrices to map exactly which neural weights were most responsible for memorizing a specific data point, neutralizing only those points surgically.

Preservation by Distillation (SCRUB): uses a "teacher" model to ensure that while the target memory is being destroyed, the model's global intelligence and performance are actively protected against degradation.

Amnesia Audit (Membership Inference Attacks — MIA): our compliance validation tool. We use adversarial algorithms that try to "interrogate" the model to find out whether a specific data point was ever part of its training. If the attacker can't figure it out, we have auditable proof that the forgetting was effective.

At the end of the day, Machine Unlearning transforms an infrastructure cost problem into an agile governance solution.

LGPD and GDPR compliance doesn't have to be the bottleneck of innovation. Without unlearning techniques, maintaining governance over AI models is slow, expensive, and operationally risky.

With the architecture we're building at Neuro, removing sensitive data from a production model becomes a lightweight process that costs a fraction of a cent in compute, takes seconds instead of days, and generates statistical, auditable evidence for legal teams and regulatory agencies.

We're already expanding our pipelines to more complex datasets (like CIFAR-10 and corporate tabular data). Neuro's next big milestones include validating these techniques on Large Language Models (LLMs) and packaging this technology into an agile Unlearning-as-a-Service API.

We believe it's not enough for AI to be intelligent — it needs to be governable. We're building the tools that give companies real control, transparency, and compliance over what their models know (and, fundamentally, over what they should forget).

If your company trains models on sensitive data and needs to solve the LGPD/GDPR compliance equation without blowing up operational costs, let's talk.