Quantifying the Brain Predictivity of Artificial Neural Networks With Nonlinear Response Mapping

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Neuroscience

Background:

• Quantifying the similarity between artificial neural networks (ANNs) and biological brains is crucial for developing brain-like AI.
• Neural predictivity, assessing an ANN's ability to predict brain responses to stimuli, is a key metric in this field.

Purpose of the Study:

• To introduce a novel method for quantifying neural predictivity using non-linear mapping.
• To evaluate the impact of non-linear mapping and network sparsity on neural predictivity.
• To compare advancements in ANN classification performance with their neural predictivity.

Main Methods:

• Proposed a non-linear mapping function to relate ANN activations to brain responses.
• Utilized a neural network to approximate the non-linear mapping, trained on neural recordings.
• Implemented and evaluated the method on 8 state-of-the-art image recognition ANNs using the TensorFlow framework.

Main Results:

• Non-linear mapping significantly improved neural predictivity compared to linear methods.
• Recent improvements in ANN classification performance did not correlate with enhanced neural predictivity.
• Network pruning (increasing sparsity) was shown to positively impact neural predictivity.

Conclusions:

• Non-linear mapping is a more effective approach for quantifying neural predictivity.
• Current leading image recognition ANNs do not necessarily exhibit better brain-like properties.
• Network sparsity, achieved through pruning, enhances the neural predictivity of ANNs.