Suggestions
Ashutosh Saxena
Founder & CEO at Caspar.AI and a prominent figure in AI & Machine Learning
Ashutosh Saxena is a prominent figure in the field of artificial intelligence and robotics. Here's a summary of his background and achievements:
Education and Academic Career
Dr. Saxena received his Ph.D. in Artificial Intelligence from Stanford University, where he worked with Professor Andrew Ng.12 He later became a professor at Cornell University, where he served for 5 years from 2009 to 2014.2
Professional Experience
Current Roles::
Previous Positions::
- Co-founder of Katapult (2012-2016)2
- Founder & Director of the RoboBrain project at Stanford University (2013-2014)2
- Chief Scientist at Holopad (2010-2013)2
- Co-founder and CTO of ZunaVision (2008-2010)2
Research and Expertise
Dr. Saxena specializes in Agentic AI for Autonomous Systems, including robots, cars, and healthcare.1 His work has been widely recognized in the field of artificial intelligence and robotics:
- Published approximately 100 papers in top venues like NeurIPS, ICML & CVPR1
- Received numerous best paper awards and over 20,000 citations1
- His work on large-scale models for robotics was awarded the 10-year test of time award in 20231
Notable Projects and Achievements
- Caspar AI: A health platform for early detection of incidents and diseases2
- RoboBrain project: Described as "Wikipedia for Robots" by MIT Tech Review1
- Brain4Cars: Predictive technology for driving mistakes1
- Make3D: Technology to turn 2D pictures into 3D models1
Awards and Recognition
- Named one of "Eight Innovators to Watch in 2015" by the Smithsonian Institution2
- Received the World Technology Award in 20152
- Selected as one of MIT Technology Review's TR35 (35 Innovators under 35) in 20182
- Awarded the Google Faculty Award (2011), Alfred P. Sloan Faculty Fellowship (2011), and NSF Career Award (2013)2
Dr. Saxena's work spans various aspects of AI and robotics, from perception and manipulation to healthcare applications, showcasing his significant contributions to the field of artificial intelligence and its practical applications.
Highlights
Gen AI startup journey :-)
With multi-modal LLMs, the surface area for attacks increases significantly. Simple prompt filters no longer work, as embeddings become non-interpretable due to high dimensionality. This CVPR paper describes an attack in the context of image generation.
https://lnkd.in/gM_ycqTs
What does this mean for domains like healthcare and autonomous vehicles?
-
Healthcare applications that use reports, scans, or X-ray images as input are now vulnerable to attacks.
-
ADAS systems using VLAMs with camera inputs are susceptible to adversarial attacks.
What is the solution?
Future ideas would revolve around:
-
(short-term) designing information bottlenecks to force interpretability (with further manual prompt and outcome filters).
-
(long-term) Safety AI agents would, as part of agentic architecture, continuously adversarial-Ly help evolve the LLMs to repel the attacks.
