Pareto-Improving Releases for ML Models
It is common to release a model if it is significantly better on average than the old model on test data. But significantimprovements in metrics don’t always mean a ‘better’ answer for everyone. To address that, especially when the test set reflects a diverse set of cases, e.g., stores in different locations, etc., it is common to gate releases of models that do worseon a segment. You could extend the reasoning to only release if the loss on each observation in the test set is no higher than from using the previous model. Pareto-improving releases (~ which have first-order stochastic dominance) producecredible continuous improvement and bring ML engineering more in line with software engineering. (Minor accommodations like accepting small performance declines may prove a more sustainable path.)
We need a Blue Apron like service for puja. Send people all the key things they need for the festivals for doing puja. Propose 5-main-festival/10-main-festival bundles. I think it is a winner.
Exploiting Sophisticated Browsers to Detect Fake News (07/2019)
Companies like FB use features of the news story, e.g., WHOIS data, metadata about the IP address hosting the domain, language used in the article, etc., to predict whether or not a news article is fake. Another good signal, however, potentially remains unexploited.
Generally, companies like FB also know which people when offered fake news don’t click on it. And companies can use the propensity to click on labeled low-trust websites to build browsing sophistication scores. And then use the early behavior of sophisticated browsers to predict which new article is fake or not.
Password Predictor (2017)
Create character level low dimensional embeddings of passwords using any of the many publicly leaked password lists and combine it with user name data to predict the password. When users are setting their passwords, tell them how many tries would be needed to guess their password given their username.
Recommended Reviewers (5/2019)
Many journal publishing systems give authors the opportunity to suggest reviewers. Some journals even require it. Create a reviewer recommendation system that initially simply samples from the citation list and eventually builds more sophistication, exploiting data on reviewers who have reviewed similar manuscripts, with some constraints for reviewer load, etc.
Detect Duplicate Images (People) in Customer Testimonials (5/2019)
Given the efficacy of testimonials, many companies contend that they have glowing reviews from actual customers. But now we have some tools to detect fake customer testimonials—look for how often the same model is used across products, etc. and dock from ratings or shame appropriately.
Auto-generate Sports Commentary (12/2017)
Let’s use cricket as an example. All official international cricket matches and many heavily viewed domestic matches like IPL use a lot of technology to capture the action on the field. The action is covered by a video camera from multiple vantage points, there is ball-tracking and there are speed guns. As a result of the heavy investment in technology, we can observe the action very well. The other great thing about cricket is that a vast majority of the action falls into a few well-known categories. For instance, batsman generally pulls, cuts, sweeps, or drives a ball. With ball tracking and speed gun, we know how quickly the ball is bowled, where it pitches, and how much it deviates off the pitch, etc. So on and so forth. And if that was not enough, we have a ton of ball-by-ball commentary of lots of cricket matches available online from espncricinfo.com. In all, we have all the ingredients to build an effective ML system to auto-generate commentary. As goes for cricket, so goes for many other sports. Sell the auto-generator to people who write commentaries so that they can more easily write better commentaries for the match discussing yet more strategic, harder-to-capture, aspects of the sport.
AI to Detect Emergency Vehicles (10/16/2017)
App. that detects distant emergency vehicles through flashing lights and sound and provides alerts on the screen. Since good microphones are much more sensitive, longer-term win = lower noise pollution.
Bad Weather Discounts (5/25/2017)
When it rains, the restaurants are generally a lot emptier. There is no reason they should be as empty as they often are. An app that automatically offers discounts on rainy days (or based on weather forecasts), and that calibrates the discount based on returns is easy to build, and likely to be useful. Integrate it with UberEats or Google Local Ads and you have an ecosystem. The general idea is about a platform for offering dynamic pricing for physical retailers. The platform will do two things: a) make it a lot easier to calibrate pricing, and b) make it a lot easier to publish discounts (where it publishes and at what price can also be optimized).
Behavioral Insurance/Mortgage Company
Lower interest rate/premiums in return for CBT + Numeracy courses.
Rate the Peer Review
One problem with peer reviews is the lack of incentives except for slack from editors. Making reviews public comes with the consequences of speaking truth to power. Making only ratings of the review (elicited from peers assigned to the same paper) public allows reviewers to get credit without the downsides.
(Stop) Waiting for the Doctor (1/3/2017)
Doctors’ offices all across the US are well-known for one thing: waiting times. Nobody expects to go to a doctor’s office and not wait. And with simple technology, if not eliminate it, we can certainly reduce waiting times. Start by collecting data on schedules, arrival times of patients, wait times, and non-arrivals. Based on current scheduling, predict wait times for people scheduled for X pm on y-day at a particular office. Adjust the schedule to reduce the estimated wait time. Add frills like texting updates if some patient interaction goes for too long, mobile phone buzzers that ring with a 10-minute warning, etc., and payout every time wait for more than X minutes.
Alternately, one could create a market for trading waiting times? Most will go back richer, even though even the rich generally undervalue their time.
Obamacare leverages a ‘smoker’s surcharge.’ In the same spirit, an anti-vaxxer surcharge can be easily instituted. The latter still doesn’t account for the externalities, but still better than nothing.
Chatbot for Data Science
A well implemented AI chatbot that provides an interface to Azure + data munging and has strong visualization support would reduce the cost of data science dramatically. (12/5/2016)
Surveys in DMV
State DMVs are places where a broad cross-section of population wastes time. Thus, they are excellent places for surveying representative samples.
Female Drivers for Uber
Concern about sex crime in India once led the Delhi government to shut down Uber. To address concerns, Uber in India may want to default to disclosing the gender of the driver, allowing female riders to pick female drivers. This change may also have the virtue of increasing the number of female drivers (and employment).
Dropbox: git with it (7/2015)
Cloud storage platforms were invented to take headache and anxiety out of storage. And default storage of all saves with minimal information to distinguish versions is a reasonable implementation of the service for lay users. But Dropbox can do more for expert users. And over the longer term it may want to help lay users come up with better—still easy—workflows. One of the innovations that expert users want is better version control. Currently, it appears that time stamp and author information are the only pieces of metadata stored with each version. Allowing users to store more information with each save, and giving them the power to delete intervening saves is liable to prove useful.
Consumer Surveys As Payment
Providing alternate ways of access content is liable to be a win-win-win: users, content producers, and platforms.
Survey Companies: Exploiting Collected Data II
YouGov, like many other survey companies, conducts lots of surveys. But it fails to exploit the databank adequately. Assuming that the survey data are the intellectual property of the person (organization) who sponsors the survey, it could allow survey sponsors to enroll in a loyalty program where they get paid every time another user wants to use the data they have collected. To that end, they would need to create a searchable database of questions. Such a database is liable to prove lucrative for both YG and researchers. It is also liable to have positive externalities for research.
Survey Companies: Exploiting Collected Data I
An average panelist fills out a lot of surveys. This means that survey companies have extraordinarily rich data person. They can use that data to do better matching and to estimate measurement error. For instance, if a survey respondent fills out a question, across say 1000 other surveys, even 1 second, an average survey response time of .5 seconds on a survey may be flagged as a potential outlier or used to build an estimate of potential measurement error.
Cut from successful ventures
NIH doesn’t take a cut from successful investments. It is not clear why not. (NIH can still fund losing ventures as important areas may not always be lucrative.)
More Idealogs…keep the fire burning.