The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Skip-Convolutions for Efficient Video Processing with Amir Habibian - #496
Jun 28, 2021
In this engaging discussion, Amir Habibian, a Senior Staff Engineer Manager at Qualcomm Technologies, delves into groundbreaking advancements in video processing. He explores the innovative concept of skip convolutions, which enhance efficiency in visual neural networks. Amir also introduces his FrameExit framework, a conditional early exiting mechanism for video recognition, optimizing how frames are processed. The conversation highlights the future of AI in video technology and the importance of tailoring methods to maximize performance.
AI Snips
Chapters
Transcript
Episode notes
AI Origin Story
- Amir Habibian's interest in AI began with an online chess-like game competition during his undergraduate studies.
- This sparked his pursuit of AI, leading to a PhD in computer vision and a focus on video understanding.
Inefficient Frame Processing
- Traditional video processing treats video as a sequence of static images, processing each frame individually.
- Amir Habibian argues this is wasteful and suggests leveraging the correlations between frames for efficiency.
Activation Redundancy
- Habibian suggests that focusing on the redundancy in activations, rather than weights, is underexplored.
- This approach can significantly improve computational efficiency in video processing.