Related Experiment Video
Updated: Dec 12, 2025
12:27
Large-scale Reconstructions and Independent, Unbiased Clustering Based on Morphological Metrics to Classify Neurons in Selective Populations
Published on: February 15, 2017
7.2K
Improved object recognition using neural networks trained to mimic the brain's statistical properties.
Callie Federer1, Haoyan Xu2, Alona Fyshe2
1Department of Physiology and Biophysics, University of Colorado Anschutz Medical Campus, Aurora, CO, United States of America.
Summary
Training deep convolutional neural networks (DCNNs) with brain-like representations improved object recognition performance and robustness. This approach, using statistical properties of neural data as a teacher signal, enhanced DCNNs without direct neural data input.
Area of Science:
- Neuroscience
- Computer Science
- Artificial Intelligence
Background:
- Deep convolutional neural networks (DCNNs) are inspired by the mammalian visual system and achieve human-level object recognition.
- Existing DCNNs trained for object recognition develop representations similar to those in the mammalian visual system.
- This similarity suggests that optimizing DCNN representations for brain-like patterns could enhance performance.
Purpose of the Study:
- To investigate if training DCNNs with brain-like representations improves object recognition performance and robustness.
- To test the hypothesis that a composite training task, incorporating representational similarity to neural data, enhances DCNNs.
- To determine if direct neural data is necessary for this representational improvement.
Main Methods:
- Trained DCNNs on a composite task requiring both object classification and similarity of intermediate representations to neural recordings from monkey visual cortex.
- Compared performance and robustness of DCNNs trained on the composite task versus those trained solely on object categorization.
- Investigated the effect of using randomized data with similar statistical properties to neural data as a training signal.
Main Results:
- DCNNs trained on the composite task exhibited improved object recognition performance and greater robustness to label corruption compared to standard training.
- Performance gains were observed even when using randomized data with the statistical properties of neural data, indicating neural data itself was not essential.
- These performance improvements were consistent across various network architectures, optimizers, activation functions, and initialization methods.
Conclusions:
- Training DCNNs to emulate brain-like representations offers a viable strategy for enhancing object recognition capabilities.
- The statistical properties of neural activation patterns, rather than the raw data, can serve as an effective teacher signal for DCNN training.
- This approach demonstrates a novel method for developing more effective and robust object recognition algorithms by leveraging principles from neuroscience.
