Chapter
Challenges in Building Universal Invariance for Machine Learning
51:37 - 56:07 (04:30)
The speaker discusses the challenge of selecting a small set of admissible functions that are good enough to extract universal invariance for machine learning and image understanding, which requires several invariance to get the same performance as the best neural net using 100 times fewer examples.
Clips
The challenge in selecting a function from an infinite number of predicates that is admissible and small enough to solve a problem, can be addressed using invariance.
51:37 - 53:07 (01:29)
Summary
The challenge in selecting a function from an infinite number of predicates that is admissible and small enough to solve a problem, can be addressed using invariance. However, it requires selecting a finite set of invariance rules that mimic human understanding, as logic-based systems are not sufficient on their own.
ChapterChallenges in Building Universal Invariance for Machine Learning
EpisodeVladimir Vapnik: Predicates, Invariants, and the Essence of Intelligence
PodcastLex Fridman Podcast
The journey from solving hand-written recognition to understanding more general natural images is a challenging one, requiring the creation of several invariances that can match the performance of the best neural networks with far fewer examples.
53:07 - 56:07 (03:00)
Summary
The journey from solving hand-written recognition to understanding more general natural images is a challenging one, requiring the creation of several invariances that can match the performance of the best neural networks with far fewer examples.