Deep Neural Networks: A New Framework for Modeling Biological Vision and Brain Information Processing

Area of Science:

• Artificial intelligence
• Computational neuroscience
• Computer vision

Background:

• Neural network modeling has significantly advanced artificial intelligence (AI) and computer vision.
• Artificial systems are approaching human-level visual recognition capabilities.
• Artificial neural networks are inspired by biological brains and their computations.

Purpose of the Study:

• To explore the potential of artificial neural networks in modeling brain computations.
• To investigate the biological faithfulness of current artificial neural network models.
• To pave the way for building biologically accurate computational models of intelligence.

Main Methods:

• Leveraging advances in neural network modeling, particularly convolutional feedforward networks.
• Drawing inspiration from the primate visual hierarchy for network architecture.
• Comparing internal representations of artificial models with primate brain data.

Main Results:

• Artificial systems are nearing human-level performance in visual recognition tasks.
• Initial comparisons reveal striking similarities between the representational spaces of ANNs and primate brains.
• Current models, while engineered, show promise for biological modeling.

Conclusions:

• A new era of building biologically faithful computational models of intelligence is emerging.
• These models can capture high-level cognitive functions like vision.
• Future work will focus on both feedforward and recurrent neural network architectures for greater biological accuracy.