Researchers at ICTP Trieste have unveiled the theoretical mechanism behind how deep neural networks learn from data, promising more efficient AI systems.

For years, the inner workings of deep neural networks have been treated as a black box. While engineers knew how to build them and data scientists knew how to feed them, the precise theoretical mechanism that allows an artificial intelligence to extract meaningful patterns from a sea of raw data remained an elusive mystery. This gap in understanding has persisted for decades, leaving the industry to rely more on empirical trial-and-error than on a fundamental scientific blueprint.

A breakthrough has now emerged from Trieste, Italy. A team of five researchers at the International Centre for Theoretical Physics (ICTP) Abdus Salam, led by Jean Barbier, has successfully developed a theoretical description of the learning process within a specific class of deep neural networks. Their findings, published in the prestigious journal Physical Review X, provide the missing link in understanding how AI identifies the most relevant information within complex datasets.

Decoding the mechanics of feature learning

At the heart of this discovery is the concept of feature learning. This is the specific ability of a neural network to identify hidden structures in initial data and generalize that knowledge to new, unseen information. Without effective feature learning, an AI would simply memorize a dataset rather than understanding the underlying logic, rendering it useless for real-world applications.

The ICTP team focused their analysis on the regime of near-interpolation. This is a critical state where the number of parameters within the neural network is roughly comparable to the volume of data used for training. It is precisely in this balance that authentic feature learning occurs. By isolating this condition, Barbier and his colleagues were able to describe how the network transitions from simply processing data to actually learning the essential characteristics that define a pattern.

A fusion of physics and mathematics

To solve a problem that had blocked the scientific community for years, the researchers stepped outside the traditional boundaries of computer science. They combined two sophisticated theoretical frameworks from the world of physics to map the behavior of AI.

First, they utilized the theory of disordered systems, a tool used by physicists for over forty years to analyze how patterns emerge and become detectable within complex, chaotic data. Second, they integrated random matrix theory, which is essential for describing the statistical behavior of large-scale systems. By merging these two approaches, the team created a mathematical lens capable of observing the invisible processes that occur as a deep neural network optimizes its weights during training.

The theoretical framework we have developed is the key that was missing to understand how the complex neural networks at the base of today's applications operate.

Reducing the cost of intelligence

The implications of this research extend far beyond academic curiosity. For entrepreneurs and tech leaders, the primary hurdle in scaling AI is the astronomical cost of compute and energy. Current training methods are often inefficient because they rely on brute force—throwing more data and more GPUs at a problem until it works.

By understanding the fundamental laws of how AI learns, developers can now move toward designing algorithms that are inherently more efficient. This means training models that require less data to reach the same level of accuracy and reducing the computational overhead. In a market where the cost of training a frontier model can reach hundreds of millions of dollars, a theoretical roadmap for optimization is a significant economic lever.

The broader European digital push

This scientific milestone in Trieste arrives amidst a wider strategic effort to secure digital sovereignty in Europe. The ability to optimize AI at a theoretical level complements massive infrastructure investments currently underway on the continent. For instance, the Schwarz Group has announced plans to invest up to 5.6 billion euros by 2033 in a data center in Mecklenburg-Vorpommern, Germany. This project aims to foster cloud and AI solutions operating under European law.

The synergy between theoretical breakthroughs, like those from the ICTP, and infrastructure investments suggests a dual-track strategy: building the physical capacity to host AI while simultaneously refining the mathematical efficiency of the software running on those servers. This approach reduces dependence on non-European proprietary architectures and lowers the energy footprint of the digital transition.

Global business impact and regulatory outlook

For international firms, particularly those in the USA and UK, the ICTP discovery signals a shift toward a more transparent era of AI. The transition from empirical AI to theoretical AI will likely influence how companies approach model auditing and safety.

In the United States, where the regulatory environment remains fragmented and focused on voluntary commitments and executive orders, a better theoretical understanding of AI learning could provide the basis for more objective safety benchmarks. Instead of relying on 'red-teaming' (trying to break a model to find flaws), developers may eventually be able to mathematically prove certain properties of a model's learning process.

In the UK, which has positioned itself as a hub for AI safety, this research provides a tool for the AI Safety Institute to better analyze the risks associated with emergent properties in large models. If the mechanism of feature learning is fully understood, it becomes easier to predict when a model might develop unintended or dangerous capabilities.

Furthermore, while the EU AI Act focuses heavily on risk categories and transparency, the ability to optimize algorithms for efficiency aligns with the EU's broader Green Deal goals. Reducing the energy required for AI training is no longer just a cost-saving measure; it is becoming a regulatory necessity as carbon reporting for data centers becomes more stringent across the global market.

FAQ

---

Sources: Gazzettadiparma, It, Nordestnews ·