A Pitch for �High-Concept Research

Noah Snavely

Cornell Tech & Google DeepMind

“High concept”

• Hollywood term that is a little hard to define precisely (ala mise-en-scène)
• Wikipedia: “A type of artistic work that can be easily pitched with a succinctly stated premise” (or logline)
• Premise that is unique, easily explainable, and instantly engaging
• “…typically characterized by an overarching ‘what if?’ scenario that catalyzes the following events”
• What if toys came to life when humans weren’t around?

Examples – film

Scientists clone dinosaurs and create a theme park, but things go wrong when the dinosaurs escape and wreak havoc.

Examples – film

A cop must prevent a bomb from exploding on a city bus by keeping its speed above 50 mph.

Examples – film

A skilled thief who can enter people's dreams and steal their secrets must perform the ultimate heist by planting an idea into a target's subconscious.

Low concept

• When a reporter is assigned to decipher a newspaper magnate’s dying words, his investigation gradually reveals the fascinating portrait of a complex man who rose from obscurity to staggering heights.
• A young man and woman meet on a train in Europe, and wind up spending one evening together in Vienna.

​

• Not a judgment of bad or good, just different types of stories

High-concept research

• A research premise that is “sticky”, and where once you hear it, you are eager to start mentally filling in some of the details in your head
• Fun, appealing ideas
• Some of my all-time favorite papers are in this category
• A (very incomplete) set of examples follows

Examples

Can a person be a camera?

Accidental pinholes and pinspecks [CVPR12]

Accidental pinholes and pinspecks [CVPR12]

Accidental pinholes and pinspecks [CVPR12]

Accidental pinholes and pinspecks [CVPR12]

Examples

Can a person be a light source?

Humans as Light Bulbs:

3D Human Reconstruction from Thermal Reflection

Ruoshi Liu, Carl Vondrick

CVPR 2023

Can an X be a Y?

CVPR 2005

A Century of Portraits

Pioneering paper in visual discovery

• c.f. Doersch, et al. What Makes Paris Look Like Paris. SIGGRAPH 2012
• c.f. Gebru, et al. Using deep learning and Google Street View to estimate the demographic makeup of neighborhoods across the United States. PNAS 2017

CVPR 2010

Depth Map

High-concept research – Cons

• Might not get 10 million citations; might be “research islands”
• Doesn’t obviously advance the main lines of research in the world
• Digging up some weird fossilized animal skeleton vs. mining for gold
• Doesn’t explore the “deepest” scientific questions from first principles
• Can be hard to convince others it is a good idea to work on
• Can be very hard to come up with good high-concept ideas

High-concept research – Pros

• Research can stand out as clearly distinctive
• Can make for a good story, oral presentation, job talk
• Might just get an award
• Can tend to be evergreen ideas – might be okay to wait years to work on
• Inspiring, might encourage others to get into a specific area of research (or get into research in the first place)
• Might keep people up at night wondering “how can I apply that idea in some other way”?
• Fun to work on
• Caveat: many examples I’ve given are from established groups

Free idea: Hidden rainbows

No Rainbow, Then Faint Rainbow, Then Rainbow

Hypothesis: there are photos of the sky with rainbows that are just beyond the limits of human perception. Let’s find them!

Takeaways

• Take all takeaways with a grain of salt (survivorship bias, etc.)
• Invest time in finding good, interesting problems
• It’s okay to not try to solve intelligence
• It’s okay to find and work on problems that you think are fun
• Please work on finding hidden rainbows!

Thank you!