Lots of companies use A/B test results as a way to measure the value of features to their users.
Do you think the practice is still relevant? I feel that A/B testing needs to be phased out by an infrastructure leveraging ideas from Contextual Bandits, Deep Reinforcement Learning and Counterfactual Policy Estimation.
The features would get faster to the market and companies would be able to use the best algorithms for a given context.
What are your thoughts?
Thursday, January 30, 2020
Saturday, September 30, 2017
Neural Dialogue Generation
Impressed by authors of Adversarial Learning for Neural Dialogue Generation (https://lnkd.in/g3rd3Zf). They have creatively combined Reinforcement Learning, Sequence to Sequence Model and Adversarial Network to generate human like Natural Language Conversations.
Thursday, September 28, 2017
Thoughts on Seq2Sql research paper
Spent this week reading the research paper (https://lnkd.in/gcrqPnR) from Salesforce Einstein team. I was impressed to know that Seq2sql model got 70% accuracy in generating SQL from Natural Language. The authors have exploited structure of the SQL and creatively combined LSTM'S and Reinforcement Learning approaches
It would be great to know Linkedin community's thoughts on how we can further increase the accuracy of SQL generation? What are the approaches to extend the problem to generate Join Operations and Multi Step SQL statements. How far do you think, we are from replacing Data Analyst's with smart enterprise Information Virtual Agents?
Catching up with developments in Recommendation Algorithms using Deep Learning
In my quest to identify technology and business gaps for Voicy.AI, I have been spending time to catch up with developments in Recommendation Algorithms using Deep Learning. I started my research reading Recommendations paper from YouTube. Recommendation in general is a two step process consisting of Retrieval and Re-Ranking. The authors have phrased the retrieval as a Multi Class classification instead of reusing Inverted Index scoring mechanisms. I liked the tricks of Negative Sampling and Sub-Linear scoring using hashing techniques to optimize for training and serving in production respectively. I than moved to another important development in the recommendation systems leveraging joint training of Wide And Deep Learning Neural Nets pioneered by Apps team at Google. I was impressed with the observation of authors about how Wide Model is good for Memorization and Deep Model is good for generalization.
I than stumbled upon another research paper from UCL folks. The authors have focused on the retrieval problem of recommendations in the context of journalism. I liked how the authors have used structure of the problem and seperate attention models to construct profiles to predict recommendations. It was impressive to see the big leaps by DL algorithms for Recommendation problem from Collaborative Filtering algorithms few years back.
What is your opinion about the next DL paradigm for recommendations? Any suggestions for more popular research papers in DL based recommendations?
Thursday, April 9, 2015
Rambling about correlation from an insight in movie to ab test
Given that I work on problems to improve relevance for Amazon Instant Videos, I have been watching amazing videos on amazing Amazon :).
I use the pretext of "learning" from different situations in movies to justify my conscience to spend time on watching videos vs reading books(Work in progress to get out of this justification :)). Based on a recommendation from my manager at Amazon, I started watching Vikings series. One sentence, in the series, which I thought was insightful was "Never fight a war, unless you know that you are going to win". The movie dialogue kept me thinking more. How do you know, what are your odds of winning without fighting a war ?.
History was made in battles where the odds of winning were lower, because of great leaders. Alexander, won over a much bigger Persian army with much smaller army. Babar, the Mughal empire founder won over a much bigger army. Do you attribute the wins to leaders or situations or armies ?. Does anyone fight a war when you know that you are going to loose ?. As a leader, how do you rationalize that you can win, despite the evidence of having less numbers or not having know how ?.
I think the decision making process of predicting winning a war has become easier over the generations. Let us say that you are planning a ab test to improve customer experience for movie recommendations how do you evaluate if your test is going to win before releasing it ? You can use anecdotal evidence, look at customer studies, run offline evaluations and get a direction sense of the weblab success. Even after doing all the diligence you still fail, specially when you are optimizing on multiple dimensions. How do you know you are getting close to that pareto optimal point in your weblab ?. Yes, You can potentially plot data offline and evaluate it, but you might still fail :).
I use the pretext of "learning" from different situations in movies to justify my conscience to spend time on watching videos vs reading books(Work in progress to get out of this justification :)). Based on a recommendation from my manager at Amazon, I started watching Vikings series. One sentence, in the series, which I thought was insightful was "Never fight a war, unless you know that you are going to win". The movie dialogue kept me thinking more. How do you know, what are your odds of winning without fighting a war ?.
History was made in battles where the odds of winning were lower, because of great leaders. Alexander, won over a much bigger Persian army with much smaller army. Babar, the Mughal empire founder won over a much bigger army. Do you attribute the wins to leaders or situations or armies ?. Does anyone fight a war when you know that you are going to loose ?. As a leader, how do you rationalize that you can win, despite the evidence of having less numbers or not having know how ?.
I think the decision making process of predicting winning a war has become easier over the generations. Let us say that you are planning a ab test to improve customer experience for movie recommendations how do you evaluate if your test is going to win before releasing it ? You can use anecdotal evidence, look at customer studies, run offline evaluations and get a direction sense of the weblab success. Even after doing all the diligence you still fail, specially when you are optimizing on multiple dimensions. How do you know you are getting close to that pareto optimal point in your weblab ?. Yes, You can potentially plot data offline and evaluate it, but you might still fail :).
Wednesday, December 24, 2014
Virtual Advertising in Videos
For one of the hacks at A9/Amazon, I teamed up with one of my co-workers to insert ads into videos for brands. The idea was to identify spots in image frames where you can embed brand ads to present a seamless experience to the users. We used opencv, choose sample videos, used machine learning, color and shape detection algorithms to figure out spots in image frames which can be swapped out with banners of ads. After embedding ads into frames we re-composed back the video. For monetization, we proposed pay for play to the advertisers.
It was a lot of learning and hacking in one day. The crowd who looked at the hack were mesmerized with the idea. After the hack presentation, we found out that one of the judges, founded a start up with the same idea. It was a lot of learning and hacking in one day. A very memorable experience.
It was a lot of learning and hacking in one day. The crowd who looked at the hack were mesmerized with the idea. After the hack presentation, we found out that one of the judges, founded a start up with the same idea. It was a lot of learning and hacking in one day. A very memorable experience.
Saturday, January 11, 2014
Excerpts from my older blog on internet archive
Using the power of web archive project, I was able to get content from my old blog, I used to publish in 2007-2008 time frame. You can see my previous blog here at http://web.archive.org/web/20080420212438/http://www.techiemusings.com/
Here are some interesting articles I picked up:
Here are some interesting articles I picked up:
Location aware Mobile Applications
We now have a OS on mobile phones exposing location as a attribute with Google’s Android. The practical applications are huge, the simplest being a poor man’s GPS if you combine with public Wifi’s. I did some research to see if there are any mobile browsers or mobile applications, exposing information about location of a user. Interestingly enough, I found information about twibble a location aware mobile client, metosphere a mobile browser for Google Android.
I have been researching around to see the progress in the industry on location based advertising. I stumbled upon proximitymedia a contextual advertising firm which uses blue-tooth network to target content/ads to the user. Applications are end less. A customer gets lunch coupons, when a user enters mall during lunch time, sees movie trailers when he is waiting in a queue for tickets.
Content Recognition Software
With the proliferation of peer to peer networks and youtube’s, I was researching to see, if I could find any software company which is cashing on copyright infrinngement detection. The software is definitely complex, given the scenario of multiple devices, mutliple wave formats and multiple encoding formats. A variation of fingerprints, hash signatures using Fast Fourier Transform are used to automate the detection of copyright infringement. Virage a division of Autonomy software is helping big name clients like BBC for detection of copyright infringement.
Social Ads and its hidden Implications
I was discussing about Social Ads from Facebook and its implications with one of the co-worker. In my previous post, I was ignoring one major advantage of a Social Network for advertisers, the “social effect”. If an advertiser is able to convert (acquire) at-least one user in a social network, the user can influence decision making of his close friends in the network and hence indirectly increase acquisitions for the advertiser. Google is doing all the right things to prevent Facebook from becomming the dominant Social Network. It would be interesting to follow the adoption of Social Ads across the web.
Social Ads from Facebook
Facebook is throwing a direct challenge at Google by invading on its domain. It is all set to release “Social Ads” a platform for outside developers to develop niche applications to target ads to facebook users outside of network, based on cookies placed by facebook. This can be a game changing play from facebook, as profiles of facebook users is a rich source of information to target ads. That partly explains the lofty market cap Microsoft gave to Facebook. Google in-turn, has released Open Social platform to dilute the effect of Facebook social network. I am sure, Yahoo is going to make moves in this domain, as it is still the world’s most popular social network.
Peer to Peer Knowledge Management
I was researching articles to know more about peer to peer networks. The recent propulsion of data transfer with an increase in youtube like sites has increased load on network. CDN’s are moving from centralized single point architectures to peer to peer nodes. This explains the recent vc funding into startups working on peer to peer cdn’s. Extending the concept, I was surprised to find the amount of research being done into peer to peer knowledge management. A direct take from one of the publications is “Of all the application domains we have studied, knowledge management is the one where the benefits of peer-to-peer and a clear and straightforward business model for suppliers are most evident”. By coupling virtual meta data of knowledge of all the nodes in a knowledge network and map of node to meta data, individudal data node can be a source of global information, making the node a better decision maker. The applications of such a framework are endless. The literature talks of knowledge management in a enterprise as a classic case
Direct Mail Targeting by Zillow
While, I was browsing through the web, I found an interesting article on how zillow is changing the game of direct mail marketing. The whole process started, sometime back in Sep 2006, when it allowed the users to add annotations to their home locations. Since, then now have more data about the adresses, and profiles of homes, zillow is selling this data to advertisers, which can then target the ads based on specific addess detail. Interesting concept. I wonder, if yahoo real estate would do anything in this space.
Story of Domain Names
Over the weekend, just for the sake of registering interesting domain names, I ran code to look for unregistered domains on yahoo domains with 3 or 4 alphabets using java httpclient. To my surprise, all the domains are allready taken. Though, I was expecting that domains with 3 alphabets being taken, but not with all the possible combinations of 4 alphabets. This effectively means, that 26*26*26*26 domains names are all ready taken. I was shocked. Yahoo with its 1.99$ domain registration deal seems to be the culprit. Another interesting conclusion, was the rate at which User Generated Content/Internet awareness is spreading over.
Silicon Valley Code Camp
I attended/volunteered for the Silicon Vally Code camp over the weekend. It was a fun event. Had a chance to talk to Peter Kellner, Dave Wright. Dave Wright was confident that it is just the matter of time, before Live.com takes over Google. The technology sessions were really great. People were talking more about commodity computing/ondemand computing, a concept started by Sun Microsystems and now mass marketed by Amazon EC2/S3 services. It is getting traction with mainstream developers and is something definite to watch about. Dave Wright was commenting that Microsoft is not going to enter into it, unless it really starts kicking in. I was also amazed by number of people interested in knowing more about SilverLight. Adobe has to watch out. Though, I completely disagree, someone in session was evangelizing upon modifying architectures to use a cache/relational db in browsers to optimize load on servers. Another interesting session was GridGain a grid computing framework for Java. Overall, it was fun 2 days.
Subscribe to:
Posts (Atom)