🧠 AI: Some architectures are fundamentally close to the human brain

Some artificial intelligence models can already resemble the human brain even before having learned anything. This surprising finding comes from a study that challenges traditional approaches to machine learning, which are often based on massive data analysis.

This research, conducted at Johns Hopkins University and published in Nature Machine Intelligence, suggests that the initial design of an AI system could be more important than lengthy training processes. The latter usually consume a lot of energy and time. The team compared several types of neural network architectures to see which produced activity patterns similar to those observed in the brain.


The researchers examined three main types of designs: transformers, fully connected networks, and convolutional networks. They created dozens of variants of these models, then exposed them to images without any training. The reactions of these systems were then compared to the brain activity recorded in humans and primates when presented with the same visual stimuli.

Among all the architectures tested, convolutional networks showed particular behavior. When the number of their artificial neurons was increased, their internal activity became more similar to human brain patterns. This capability was present from the start, without any data being used to adjust the model's parameters.

These untrained convolutional networks proved comparable to conventional AI systems, which usually require millions of images to learn. The lead researcher explained that if training on massive data was truly decisive, such results would not have been possible by modifying the architecture alone.

This observation offers new avenues for the future development of artificial intelligence. Starting with a better-designed initial architecture, inspired by biology, could significantly reduce the resources needed for learning. The work now continues to integrate simple algorithms, also inspired by living systems, into new development frameworks.

Thus, the path towards more efficient AI systems might rely less on the brute force of data and more on clever design, directly drawing inspiration from the principles proven by biological evolution.

How a convolutional network structures visual information

A convolutional neural network is organized into successive layers, somewhat like the visual cortex in our brain. Each layer specializes in detecting increasingly complex features. The first one can detect edges or color changes, while subsequent ones identify shapes or assemblies.

This hierarchical structure allows for very efficient image processing. Instead of analyzing each pixel independently, the network progressively extracts meaningful elements. It is this organization that seems to naturally provide internal activity close to that of the brain, even without prior learning.

The fact that these untrained networks already show such similarity to brain activity indicates that their 'blueprint' is fundamentally suited to the visual task. This shows the importance of carefully choosing the starting architecture, even before beginning to feed the system with data to refine it.

Why the initial design can accelerate learning

Learning in artificial intelligence often relies on adjusting millions, or even billions, of internal parameters. This process, called training, usually requires enormous amounts of examples and computing power. It generally starts from a random or very simple initial state.

If the model already starts from a state close to the solution, as this study indicates, the path to reaching a good performance level is much shorter.

This is the potential advantage of a 'cortex-aligned' architecture from the start. It places the system in a favorable configuration, reducing the need for data and iterations. This could lead to models that learn faster, with less energy, by drawing inspiration from the efficiency principles discovered by biological evolution.