How Computers Misunderstand the World

AI is a branch of computer sicence, composed of many subfields. Machine learning happens to be one of those fields. Machine Learning is a misleading term becaues there are no little brains in the machine. Machine Learning is computational statistics on steroids. Speaker likes the CMU definition by Tom M Mitchell: learning does not include intelligence. Titantic dataset will tell you the best predictor is the class they were traveling; reason was because the cooridors were locked. The data can tell us who might surivive however it misses the reason.

The data fails to capture key quantative knolls, such as the angle someone chose to dive off the Titanic. “Not everything that counts can be counted”. “Technochauvinism”: The idea the computer is always the right tool for the job. When we say computers are better than people, we really mean math is better than people.

A lack of diversity on the teams results in unexpected failures to deliver intended results. Positive Asymmetry? People are less likely to speak up and talk about issues like the racist soap dispenser. Contemporary image recongition algorithms are extremely fragile and face issues like reconginizing people with varying skin color. “AI will not deliver us from ourselves.”

Challenge from speaker: Differentiate automation and AI. “Ghost Work”: The people hiding inside of AI systems -> when the AI unit fails the work is sent to a digital sweatshop to process the work when the computer fails. Speaker challenges us to acknowledge the humans in the loop instead of pretending there are no humans or replacing all of the people.

Can Androids concieve of Electric Sheep?

Speaking about his artistic expressions. Been programming since high school days when learning C++. Working with a lot of different software. Rigging is the process of taking a model and designing an interconnected skelaton so you can pose it in multiple positions. “Imagine Bob Ross finish this before you finish your meeting”. Inverse Kinsimatics, the way they work with robots to figure out how to move an arm into the correct position. Instance segmentation is responsible for figuring out what pixels are related to which object.

Generative Art - Art programmed using a computer that intentionally introduces a value of randomness to create additional value. “Deeply artificial trees” 2017: A hallucinogenic nightmare of Bob Ross. “Creativity is just connecting things” - Steve Jobs.

Mask R-CNN trained on COCO - Common objects in context. ConceptNET matches objectst within a context.


  • OpenCV
  • GPUImage
  • Mask R-CNN (Matterport’s)
  • Unity
  • ConceptNet
  • COCO - FB, MS, etc

Correctness Proofs of dsitrubted systems with Isabelle/HOL

Testing is the answer to know the program is doing what it’s suppose to do in the traditional approach. Model checking has gained popularity, however there is an exponential state explosion due the non-deterministic nature of the systems. Formal proofs are the only way to ensure all properties are the only way to be entirely sure. Proofs have a large level of effort so they should be used for algorithms where correctness is not obvious, complicate, or might be desired to be studied.

Isabelle does not have native support for distributed systems. Agreement property: the two entities reach the same value of time. Modeled Paxos as a linear sequence of timesteps where each actor handles an event.

Rhapsody in Zero Knowledge

Speaker is a Cofounder of Iqlusion.

Problem: Identity Theft -> Data Breaches as a result of insecure authetnication mechanism

Solutions: Strong authentication mechanism -> Frightening privacy implications. Integral by the Pet Shop Boys. “We” by Yevgeny Zamyatin is the prototype for the modern dystopia. Telephone is broadcasting IMSI, Mac, & Bluetooth ID. Zero knowledge proofs offers privacy without sacraficying security.

Zero knowledge proof: Pulls from a proof of knowledge. Claim + Witness. Is a proof of a witness without any knolwedge of the witness. 1982 - Shafi was working on probablistic encryption: Trapdoor Function (pubkey) should be replaced with this type of encryption.

Watch the Watchers


face-api.js - Used for autodetect faces Lucy Parsons Labs - A collaboration between data science and oversight Fewer than 2% of the complaints resulted to any sort of results, most suspended for 1 week or less. FOIA: Warden v Police Department -> Ruled against CPD for providing photos

There is a history of Police retaliating against misconduct reporters.

Uses cropper.js to help manage and crop images.

A Mulching Proposal: A tounge and cheek exploration of the implications of using technology which improperly identifies details.

Github justadudewhohacks SSDMobilenetv1 finds a box for a face. Face Landmarks produces descriptors