Deep Networks Can Resemble Human Feed-forward Vision in Invariant Object Recognition

Area of Science:

• Computer Vision
• Cognitive Science
• Neuroscience

Background:

• Deep convolutional neural networks (DCNNs) mimic human visual processing with hierarchical feature extraction.
• Previous studies have not fully explored DCNN performance against human object recognition, especially concerning viewpoint variations.
• The similarity in architecture between DCNNs and the human visual system warrants investigation into functional parallels.

Purpose of the Study:

• To compare the performance of state-of-the-art DCNNs, HMAX, and shallow models against human object recognition under varying viewpoint conditions.
• To determine if DCNNs exhibit human-like error patterns and representations.
• To assess the impact of viewpoint variation magnitude on model and human performance.

Main Methods:

• Benchmarking eight DCNNs, the HMAX model, and a shallow model against human performance using backward masking.
• Systematically controlling the magnitude of viewpoint variations in object recognition tasks.
• Analyzing performance metrics, error distributions, and representational consistency with human behavior.

Main Results:

• Shallow networks outperformed deep networks and humans at low viewpoint variations.
• Deeper networks were required to match human performance and error distributions with increasing viewpoint variations.
• An 18-layer DCNN surpassed human performance and demonstrated the most human-like representations at high variation levels.

Conclusions:

• The depth of neural networks is crucial for achieving human-level view-invariant object recognition, particularly with significant viewpoint changes.
• DCNNs can replicate human-like representations and error patterns, especially deeper architectures.
• Network depth is a key factor in bridging the gap between artificial and human visual recognition capabilities.