Почему виртуальная муха не настоящий мозг | Владимир Алипов

The talk begins with Alipov describing a major Nature paper that built upon existing Drosophila connectome data — not new experimental data — to create a computational brain model. The researchers mapped the fly's neural connections and tested the model by sending sensory signals (like sugar flavor on one side) and observing which motor centers activated, comparing results between the live fly and the simulation. The model was placed inside a virtual environment with simulated physics, smells, and dust, where the simulated fly demonstrated basic behaviors: grooming its antennae and seeking out banana pieces to eat.

Alipov then pivots to a critical analysis of what this model actually represents versus how it has been portrayed in media. He identifies the absence of synaptic plasticity as the most fundamental flaw — the model is a static structure that cannot learn or adapt. Since neuroscience still lacks a complete understanding of how synaptic connections change during learning (long-term potentiation, immediate early genes, etc.), these processes simply cannot be modeled accurately yet.

He further criticizes the simplified neuron model used (a leaky integrate-and-fire system rather than detailed multi-compartment morphological neurons), the lack of dendritic computation, and a major shortcut in which the intermediate neural processing between sensory and motor neurons is replaced by statistical predictions. Visual processing is especially noted as completely non-functional — the fly's visual lobes are present but have zero influence on behavior in the simulation.

The result is that the simulated fly's behavior is entirely stereotypical and repetitive, lacking the natural variability and individuality of real fruit flies, which clean themselves differently, fly at different speeds, and respond dynamically to their environment. Alipov compares the gap between a real Drosophila and this simulation to the gap between a real person and a Sims character.

Despite these criticisms, Alipov concludes that the work is not meaningless — it successfully tests the hypothesis that connectome-based models can produce some emergent behavior without prior training. He acknowledges the researchers are transparent about their limitations and frames the project as a step toward understanding what works and what needs improvement, with the next stated goal being a virtual mouse brain model.