XDream

XDream discovers preferred visual stimuli for neurons in non-differentiable visual systems, such as the macaque visual cortex, by combining generative neural networks and genetic algorithms to perform activation maximization.

Key Features:

• Generative Neural Network Integration: XDream employs a generative neural network together with a genetic algorithm in a closed-loop system to create stimuli tailored to activate specific neurons, including neurons in the macaque visual cortex.
• Activation Maximization: It performs activation maximization to identify preferred features of visual units without requiring prior knowledge of their tuning.
• Comparative Efficiency: It outperforms brute-force search and exhaustive sampling of over one million images and generalizes across neurons, processing stages, layers, architectures, and developmental regimes.
• Robustness and Flexibility: Performance is robust across choices of image generators, optimization algorithms, and hyperparameters, indicating locally near-optimal solutions.
• Parameter Independence: The method does not require problem-specific parameter tuning to achieve optimal results.

Scientific Applications:

• Neural coding investigations: XDream uncovers neuronal tuning preferences to support studies of neural coding in biological preparations.
• Characterization of functional properties: It provides insights into the functional properties and stimulus selectivity of neurons in visual cortex.
• In silico experiments with ConvNet units: Researchers can use ConvNet units as surrogate models to run experiments that would be prohibitive with biological neurons.

Methodology:

XDream combines a generative neural network and a genetic algorithm in a closed-loop activation-maximization framework applied to biological neurons or ConvNet units and is benchmarked against brute-force exhaustive image sampling of over one million images.