The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Engineering a Less Artificial Intelligence with Andreas Tolias - #379
May 28, 2020
Andreas Tolias, a Professor of Neuroscience at Baylor College of Medicine, dives into the intriguing relationship between brain function and AI. He discusses how traditional AI has limitations that neuroscience can help overcome. The conversation highlights innovative data collection methods, like fMRI and high-density recordings. Tolias emphasizes the significance of understanding behavior outside neural structures and how using biological insights can enhance machine learning models. He also proposes new benchmarks to refine AI capabilities, aiming for systems that better mimic human cognition.
AI Snips
Chapters
Transcript
Episode notes
Adversarial Attacks on ImageNet
- Deep learning excels in object recognition with ImageNet, sometimes outperforming humans.
- Adding noise to images, imperceptible to humans, confuses these networks, revealing a qualitative difference in processing.
Qualitative Differences in Processing
- Deep learning models are easily fooled by adversarial attacks, highlighting their different approach compared to human perception.
- These models struggle with unknown unknowns, unlike humans and animals, because their training defines concepts differently.
Generalization and Flexibility
- While similarities exist between human and deep learning representations, key differences make networks vulnerable.
- Biological brains generalize better outside training distributions, handling new situations more flexibly.