MIT Applies Neuroscience in Efforts to Overcome Artificial Intelligence Bias


SOURCE: PSYCHOLOGYTODAY.COM
FEB 24, 2022

KEY POINTS

  • Artificial intelligence is a field of science that seeks to enable machines to perform tasks associated with the human brain.
  • When data used to train AI algorithms contain bias, results may be skewed unfavorably.
  • The ability for convolutional neural networks to generalize correlates with how specialized the neurons are.
Alexandra_Koch/Pixabay

Source: Alexandra_Koch/Pixabay

Artificial intelligence (AI) machine learning has an Achilles’ heel—if the data used to train the algorithms contain biased data, which may skew results unfavorably. In efforts to understand how machine learning can overcome this vulnerability, researchers at the Massachusetts Institute of Technology (MIT) applied the tools of neuroscience to AI and published their findings in the journal Nature Machine Intelligence.

A Weakness of Artificial Intelligence

Artificial intelligence is the field of computer science that seeks to enable machines to perform tasks associated with the human brain, such as cognition, learning, and problem-solving. For machine learning to work, there must be data to train the algorithms—massive amounts of data. The quality of AI deep learning depends not only on the algorithm but also on the data. The quality of the data set used to train machine learning can bias the algorithm’s results.

What’s not clearly understood is exactly how to overcome this weakness. In this new MIT study, researchers set out to investigate when and how convolutional neural networks fail to generalize out-of-distribution combinations that they were not exposed to during training. Convolutional neural networks are a type of neural networks used to identify and analyze images—computer vision. Biological vision has inspired AI machine learning vision.

A Neuroscience Approach

For this study, the researchers used neuroscience to help solve AI’s bias problem with data sets. In neuroscience, experiments often use carefully controlled data sets. Borrowing from this approach, the MIT researchers trained an artificial neural network for image classification with well-curated data sets to understand how well the algorithm performs when presented with out-of-distribution data that were not represented in the training data. The artificial neural networks were trained to classify category and viewpoint at the same time.

The researchers discovered that increasing data diversity while keeping a constant data set size improves the algorithm’s performance for out-of-distribution combinations but decreases the performance for in-distribution data.

They also found that separate architectures where no computing layers are shared outperformed shared ones in generalizing out-of-distribution combinations. The researchers attribute this gain to an increase in specialization. Specifically, the scientists propose the neural mechanism of specialization to “the emergence of two types of neurons—neurons selective to category and invariant to viewpoint, and vice versa.” If an artificial neural network is trained to perform tasks simultaneously versus separately, some neurons do not specialize for one task and are more likely to underperform.

Simply put, the researchers demonstrated that the ability for convolutional neural networks to generalize correlates with how specialized the neurons are. According to the researchers, their results are consistent across multiple convolutional neural networks and data sets that include the iLab-2M, MNIST-Position, MNIST-Scale, and Biased-Cars.

With these new discoveries, as next steps, the MIT researchers hope to apply their neuroscience-inspired approach to tasks with more complexity and find a way to induce artificial neural networks to develop specialized neurons in the future.

Copyright © 2022 Cami Rosso All rights reserved.