Unsupervised neural network models of the ventral visual stream.

Area of Science:

• Computational neuroscience
• Machine learning
• Primate vision

Background:

• Deep neural networks (DNNs) excel at modeling primate ventral visual stream responses.
• Supervised DNNs require extensive labels, limiting their plausibility for developmental modeling.
• Unsupervised learning advancements offer a potential solution to the data requirements for developmental models.

Purpose of the Study:

• To evaluate unsupervised learning methods as a biologically plausible model for primate ventral stream development.
• To compare the performance of unsupervised and supervised models in predicting neural responses.
• To investigate the neuroanatomical consistency and brain-like representation capabilities of unsupervised models.

Main Methods:

• Deep unsupervised contrastive embedding methods were applied to train neural network models.
• Models were trained using large-scale, real-world human child developmental data from head-mounted cameras.
• Semi-supervised approaches were explored by incorporating small amounts of labeled data.

Main Results:

• Unsupervised models achieved neural prediction accuracy comparable to or exceeding state-of-the-art supervised methods across ventral visual areas.
• The hidden layer mappings of unsupervised models demonstrated neuroanatomical consistency throughout the ventral stream.
• Models trained on noisy, limited developmental datasets produced brain-like representations.

Conclusions:

• Unsupervised learning, particularly deep contrastive embeddings, provides a biologically plausible computational model for primate sensory learning.
• These methods can effectively model neural development without extensive labeled data.
• Semi-supervised learning further enhances model accuracy and behavioral consistency.