Islands Of Intelligence

Liberating research.

2012-01-15T03:55:00.000-08:00

Identify all top research papers that deserve to be liberated by

Looking at reviews
Looking at bibliographies in text books.
Nomination by researchers
Karma donations by grad students

A parallel open paper for each paper shall be created through

community participation
homework assignments by profs

The papers shall be semantically structured for maximum reuse.
As the semantic papers suited for tablets shall be a lot more convenient than the actual papers, researchers will publish to physical journals but read semantic journals. Slowly they will also start publishing to the semantic journals.

Blitz

2012-01-15T03:14:00.000-08:00

Getting something started is usually hard.

Sometimes one needs info/understanding
Sometimes procurement is a problem
Sometimes the lack of a social network is a problem.
Sometimes some minimal labor to get something going is a problem

Groups of people could blitz a problem to get it stated. For Eg some one has a research problem and he needs a fabbed and assembled and tested PCB. Any researcher should be able to go from design to working prototype with near zero effort.

To earn his PCB he can earn karma by doing something for the community. Eg. Mashing up research papers into semantic resources. The guy who does the PCB soldering will get components for his projects. Charitable donors will sponsor the components. Donating to a blitz shall be a very high leverage donation.

Similar activities could be setting up computer vision, web scraping, google apps/aws apps for data entry and so on. It could be getting laser cut parts. It could be creating 3D models from sketches. It could be writing arduino code, it could be doing dissections remotely.... All artifacts will be opensourced as much as possible.

The basic motto is reduce the setup time as much as possible. It shall be based on a pay it forward motto. While doing so we should also reduce the total effort put forth by each individual else there is no point.

Man or Machine Game

2012-01-14T23:52:00.000-08:00

This is a proposal to teach computers and humans using a turing test as a tool.

2 randomly paired users are shown a picture
User1 asks a Multiple choice question related to the pic.
The question is shown to User2 and watson or another bot. Their answers are shown back to the asker without telling which answer belongs to whom.
User1 has to guess which is the computer and which is the person. If he guesses right he gets points.
You get khan academy like streaks for extracting different answers from the computer and the human.

In one type of implementation, the user2 is a student studying in a developing country. So it becomes really hard to distinguish between man and machine due to language and community differences. The student gets broadband credits for answering correctly. The askers could get

Game credits
Charity credits
Research credits

The answerers get knowledge and cash

a duolingo like system can be used to translate questions so that they become easier to parse for machines.

Services like Watson, Siri etc can participate as the machines. This will also serve as advertising to them. The entire knowledge base will subsequently be open sourced.

Tools for Researchers from Konstrui

2012-01-11T22:44:00.000-08:00

Thingipedia: A thingiverse for researchers. Researchers can download and build instruments or tools they need. Or order them made. It will be connected to papers that use it. Suppliers that provide materials and so on. The whole process will be transparent. Most drawings/schematics and code will be open source. The things here will make extensive use of laser cutting and 3D printing.
Rappstore: Resercher's appstore: Sudo apt-get install distributed apps on multiple devices simultaneously and get them auto configured to act like a web of devices.

Multilingual Web of Tables

2012-01-11T22:08:00.000-08:00

The web of tables is a valuable asset. To speed up the creation of the web of tables we can use Human powered translation by duolingo. This post discusses how.

Shortlist web pages with tables from

Google translate requests. ordered by no of requests. Descending.
Social bookmarking sites and social network sharing s.

Identify relational tables using strategies like

Unexpected Results in Automatic List Extraction on the Web
HyLiEn: A Hybrid Approach to General List Extraction on the Web
Entity relation discovery from web tables and links

Use a service like duolingo to translate the cells and headers. Also use heuristics to guess better translation like the values of the column must be compatible to the title.
The translation lookups will make services like google translate more valuable.
Tables in non english languages like Chinese, Japanese, Latin Americal languages, Euoropean languages etc will now be available in English and vice versa. This will unlock a lot of value.
Machine learning can be used to crawl this web of translations to infer meaning and create easy integration into tools like google refine.
Databases like Freebase and umbel could be used as a backbone for entitiy recognition

Web of Webs

2012-01-10T06:59:00.000-08:00

Web of causes - A generalisation on the lines of http://kill-or-cure.heroku.com/
Web of correlations
Web of intentional actions

W did X to Y (to Z) (because R)

Biogeochemical cycles. All flows measured in moles
Careers
Paths

Any GPS trace real or deduced.

A web of webs XPrize every year for the awesomest graph open sourced in the previous year.

Web of tables

2012-01-08T07:49:00.000-08:00

On the web tables are used for

Content
Layout
Navigation

The content tables can be identified using various pattern matching strategies and manual turking. Galleries can also be considered tables if you unwind the sawtooth into a list of cells and extract the properties

Usually content tables will have a primary key. So the mission of the game would be spotting the primary key. In some cases like [en.wikipedia.org/wiki/List_of_countries_by_population] the first column is the primary key.

Lists are in many cases just single column tables with implied columns. Eg list of Indians means, an implied column Country with value India

In some cases the top row also has primary keys. Comparison b/w objects

In some cases the top row and the first column are primary keys: Eg punnett square,adjacency matrix etc

Sometimes tables have composite primary keys. The tables can be connected to similar tables by using fuzzy mapping between attribute values. The keys can be further matched using turking. The tables can be translated into other languages using game mechanics as described earlier

The semantics of the table attributes can be extracted using games and machine learning.

A web of Reasearch tables can be created as part of Konstrui

Papers

Unexpected Results in Automatic List Extraction on the Web
HyLiEn: A Hybrid Approach to General List Extraction on the Web

Polydocs

2012-01-08T06:12:00.000-08:00

Consider a simple markup for storing data in multiple syntaxes within a single file. We already have such abilities within HTML for Eg JAVASCRIPT/CSS embedded within tags. We can cheat further and embed lanugages like Coffeescript/Handlebars templates. Now imagine a very simple text file format that stores this:

Eg:

###js url:that.specifies.grammar.or.parser url:that.provides.documentation
###sql  url:that.specifies.grammar.or.parser url:that.provides.documentation
###fly url:that.specifies.grammar.or.parser url:that.provides.documentation
###textile  url:that.specifies.grammar.or.parser url:that.provides.documentation
###latex  url:that.specifies.grammar.or.parser url:that.provides.documentation
###mwiki  url:that.specifies.grammar.or.parser url:that.provides.documentation
###reaction  url:that.specifies.grammar.or.parser url:that.provides.documentation
###uml url:that.specifies.grammar.or.parser url:that.provides.documentation
###csv url:that.specifies.grammar.or.parser url:that.provides.documentation
##js 1.js
alert("hello world");
##sql abc.sql
SELECT * FROM EMPLOYEE;
##uml dataflow.uml
A->B hi
B->A hello
B->A Bye
A->B ByeBye
##reaction #burning
2H2+O2->2H2O
##csv ["Employee","Age","Salary"] out/emp.csv
Sudu,10,10k
Budu,5,5k
Kudu,3,3k

You would load the file using a Polydoc.loadJSON(filename,optional_transformer)
It would produce a JSON object with a files attribute, which further constains a data attribute, format attribute, errors collection, args and so on. If a transformer function or script is provided, it will smush the data in the file to produce a unified DOM.

Polydocs can be stored as zip files,gists and folders

cmdline tools shall make processing easy

The "web of syntax" social network mentioned in the previous post shall be used to provide tools like section specific syntax hiligthers, section specific previwers etc. A sharejs based webapp can allow multiple parties to simultaneously edit polydocs

Web of Syntax

2012-01-08T05:52:00.000-08:00

Have a central repository of code where we have parsers for all DSLs and binary file formats that load a JSON representation of any given syntax into language JSON DOM or dumps a JSON string to stdout

SQL/XML/RDF/POV/UML/Graphviz/Assembly/RPG/SVG/STL/PCB/Modelica and so on

A method of serializing JSON to any possible format.

Web of tubes

2011-12-23T09:42:00.000-08:00

Possibly named Cloudflows : Workflows and information flows on the web.

Yahoo pipes allows you to build mashups by connecting pipes...
ROS allows you to build graphs of nodes similar to pipes.
Let us build an opensource platform that lets you connect platforms together

But just consider my brother's use case. He is a fly researcher studying neural networks in drosophila. He wants to hack together a basic solution for classifying his flies as male or female coz manual sorting is boring.

He knows that in the era of self driving cars, it should not be very hard
The algorithm itself can be a bit wasteful. There are dozens of computers in his lab staying idle most of the time. Some of them even run boinc.
His time is precious. So he does not want to search for the widest set of availability of tools. He wants to build a working system without resorting to sudo apt gets and so on.
In fact he would prefer an AMI on amazon's cloud which could be instantiated on demand.
Some day he would prefer a fully C/C++ coded MorF function to be available for the whole planet to classify trillions of drosophila. But for now, he just wants to connect his smartphone and PC over Wifi and activate an arduino controller to put the flies into 2 bins.

He builds the classifier in the browser by combining a command line and a point n click interface. The primary engine is OpenCV.

An ImageGrabber runs on the phone which sends the captured image over Wifi
to his PC running ubuntu which has ROS on it.
A background remover module removes bkgnd from all photos.
Now each photo is further purified using Mechanical Turk
An auto orientation module discovers the horizontal line
A rotator now rotates the fly so that all flies are horizontal.
Now a part spotter identifies the back region
A masker extracts only the back.
An SVM chooses b/w male and female. It sends back an M/F value.
A tiny script controls the arduino through an arduproxy module running on the phone.

Another example would be a sketchup to stl to cnc converter. I hate to create the whole pipeline. It would just be awesome to use the platform as a shortcut. In fact, I would like to start with the url of circuit diagram all the way to PCB milling. Some parts may even be turked out. A person could provide a url and produce a milling layout.

All existing tools can be chained together without much involvement by the user. For Example, raster to vector conversion, Image segmentation etc.

Some systems may even involve tele-operation using vnc. Eg u could operate sketchup or photoshop through vnc.

Everyone earns Credits by

1. Converting existing tools/web api's into tube nodes. and putting it on github/gist

2. Selling spare time on their CPUs

3. Buying Credits

4. Power turkers can do advanced stuff like circuit design to earn credits

Donations:

1. Google facebook etc can donate credits to top projects.

2. People can just invite friends to join into turkathons that help research.

The objective is to be as easy as pipes but almost any piece of useful code exisiting can be invoked. The focus is on making all code invokable easily irrespective of the platform it was written for. Focus initially on availability than efficiency. Capitalistic desires will drive up the efficiencies of the most used paths. There could be sponsored hackathons as well.

The code could be running as plugin within an app like photoshop or sketchup. or may be an independant AMI or heroku app or a firefox cloud app or what not

Users could create a warehouse api wrapping Amazon web sevices/Dropbox and so on

Companies like Google/Facebook/Sendgrid etc. could provide api credits or ad credits to top opensource contributors and as bounties for writing code given specs. These awards would also be given to those who can produce significant improvements in popular flows. Very popular flows might graduate into apis and services.

Startup academy:
As an engineer, one is trained to produce efficient, maintainable, reliable code. As an entrepreneur one has to focus on creating something that people want. To create flows on cloudflows will help u teach urself just that. It makes it easier for u to focus on what you are doing rather than how. By trying to earn credits, you have taught yourself how to code for usefulness rather than coding for the sake of coding. The cloud of code would provide a fertile ground for a lot of startups to take shape.

Especially awesome for non profits to create apps which will allow a lot of innovation.

A startup could help build an open foundation on which they can build a profitable business model. For Eg an open source platform for making 3d models from scanned images can be created while the startup tries to build up physical infrastructure.

Cloudflows and research
Researchers can quickly hack together inefficient but valuable tools to minimize drudgery. In fact we plan to dogfood our platform for building solutions at Konstrui. The problem of a lot of researchers is also the same as that of entrepreneurs. How to quickly glue together stuff to get things done.

In the world of windows we go download a huge pile of apps and do funny acrobatics to even get started. On linux we just do a sudo apt-get install everything. We want to make the experience the same for our platform.

UI for flows:

Multitouch interface to add and connect nodes. Each node will also have numbers
Cmd line where u can just type numbers like 1.2.5.4

Audience Assumption:

The target audience for the platform shall be hackers/researchers so the user interface need not be dumbed down.

Other sample applications would be an

Image/Video annotation workflow
Image or video classification or tagging
Compositing/montage/collage
Cropping
Tile generation for games
Text Annotation

Conversion bounties:
Eg :

input as xml and output as json
xslt

template language conversion
database management
query writing
code exploration: start with existing code and annotate it so that it can be rewritten.

A stackoverflow like QA forum would allow coders to collaborate and figure out how best to evolve the system.

Scaling the flows:
Each flow will have a scalability analysis where the current throughput and future optimization possibilities and savings possibilities will be shown along with effort to upgrade will be given.

Marshall Kirkpatrick's http://plexusengine.com could be a producer/consumer for modules on our plumbing...

Bootstrapping AI using duolingo: Body

2011-12-23T07:08:00.000-08:00

This is Luis von Ahn, the guy who invented reCaptcha. It has helped to OCR x million words as of XXXX. Now he has invented duolingo. He says that if a million people people play duolingo for translating English Wikipedia into Spanish, they would all have learnt English and the Spanish Wikipedia would be ready in 80 hours.

Now let us say we have 3 translations of Wikipedia, say in English, German and Spanish. Each paragraph in English would have a corresponding equivalent in the other two. Now we can use contemporary parsers like the ones used by IBM's Watson on all 3 languages. We could build parse trees for each sentence. Suppose we do a majority of 3. So that if 2 or 3 languages interpret the data in the same way, we could consider it understood. If all three parsers disagree, we could translate them into 2 more languages. If we can now get at least a 3/5th majority, we would have clearly parsed the sentences.

Once we have parse trees, we can extract, events, relationships and so on from the wikipedia. We can also do the same for Instructables, HOWTOs, Recipies, Textbooks, Howstuffworks, Video transcripts and so on. This would give us a huge treebank that stores the wisdom of our civilization in a mineable form.

I guess we can then with minimal assistance from people construct a detailed Web of production capturing supply chains, Structure of everything, Web of timelines and so on. We could even use tools like Cyc or other similar reasoning tools create these higher representations from the parse trees.

These humongous parse trees and semantic info can be used for improving parsers. Sentence templates could be extracted for better NL Generation.

Maybe the brain is just a differential equation solver.

2011-06-19T00:34:00.000-07:00

I was surprised at the simplicity of Modelica models. (https://modelica.org/) Everything that needs to be modeled is just a bunch of differential equations. Even phun/algodoo converts its stuff into differential equations and solves them. Maybe memory creation is just parameter tuning for the differential equations. Just loud thinking :)

Operational Transforms in JSON

2011-06-18T23:05:00.000-07:00

https://github.com/josephg/ShareJS

http://sharejs.org:8000/hex.html#cXZvialegs

SpaceTime Octrees are interesting and compact

2011-06-18T23:00:00.000-07:00

Papers: http://www.google.co.in/url?sa=t&source=web&cd=1&ved=0CBgQFjAA&url=http%3A%2F%2Fciteseerx.ist.psu.edu%2Fviewdoc%2Fdownload%3Fdoi%3D10.1.1.44.4687%26rep%3Drep1%26type%3Dps&ei=hZD9TYzaM8rOrQewspnDDw&usg=AFQjCNExdggQ2CcySRETdCiQmtX2tothWg&sig2=dYRwzYxQbO0ZKdzbv--lWQ

DSLs Think beyond RDF

2011-06-18T22:13:00.000-07:00

When we look at triples, it is easy get deluded into believing that Triples are a silver bullet solution to the Semantic Web. But what we really need are DSLs for every domain and mappings from those DSLs to triples. We should also be able to transform triples into DSLs. For example you can have a
1. Recipie DSL
2. Plumbing DSL
3. Geology DSL
4. Architecture DSL
5. Chemistry DSL and so on.
6. Object Structure DSL for thingiverse.
7. Web Page Structure DSL.
8. Match Score DSL.
9. PointClouds.
10. Octomap clouds.
The DSLs should be human and machine readable. An OpenSCAD model is far more useful representation of structure than a has_part relationship tree. If a has part relationship tree is needed, RDF extract can be performed on the OpenSCAD model. A GUI tool may allow a user to enter a recipie by telling. Mince(1kg of Meat)... Automatically it gets added to Ingredients.

What we miss today is a Web of DSLs. Humans end up acting as the glue between DSLs. Instead we should have programs as glue b/w DSLs. Even JPG,PNG etc are DSLs. We could even have formats like Image Annotation Language, Video Annotation Language.

The stronger the glue b/w DSLs and greater the number of transformations possible, the smarter our comps become.

More Islands of Intelligence

2009-06-17T03:53:00.000-07:00

Tim Berners Lee is making the whole world chant "raw data now"... A web of linked open data is slowly emerging...

Google has made geo data management even more indexed and easier to remix.

Wikitude combines the senses of your cellphone like GPS,motion sensors,compass etc and automagically identifies things like buildings and so on just by pointing at them through the camera of your smartphone. It can then provide contextually useful info customised to the user.

Wolfram alpha now makes anything computable just a "wolfram search away"... It is so powerful and can pull off some pretty neat tricks that a typical search engine can't. This will be an important piece of the jigsaw puzzle.

The world is slowly and steadily moving away from the "Chicken and egg" bootstrapping problem of AI.Mashing up structured data has slowly begun to produce value. Eg. Daily mile can actually sum up the number of miles you ran and add up the statistics. Imagine this, you can ask wolfram alpha to chew ur data and spew out more interesting results. The same data can flow into your profile on Facebook. These are stuff users care. But spew out data users can't even dare to imagine about. You may ask why a runner would want raw data. Coz he can use it the way he likes it. Raw data with consitent semantics and universal uris makes intelligence that much more closer.

Structured tweets can further fuel the power of the web of data. The tweets capture 'time" without need for work. The lesser the threshold of activation of creating data the more of it there will be.

And now computers can SEE!!!

How cool is that?????????????????????????????

Go read Manna. Can you dare to imagine tomorrow??
Have fun!!!

20Q to help out with object classification in a home environment.

2009-04-04T00:03:00.000-07:00

Just some dumps. Do not bother to read ;-)

Robust Object identifier.

Let us assume we need to classify any object we see at home into 1 million different classes. Cyc has about 20,000 topics common with Wikipedia. A million classes would be probably more than we care, especially assuming we have a constrained enviromne tie in and around the home. this means we need about 15 to 20 "bits" of mutually exclusive info for the classification.

1. Area of Cross section of Visible Object. Expressed in powers of 2, this would be between 45 orders of magnitude see powers of 10. In terms of powers of 2 it would be even higher. about 150 orders magnitude. But most of the objects we classify belong to a smaller range of sizes. Things smaller than buildings and larger than dust particles.
2. Density or mass can give us a range b/w tonnes and grams. But may need touching the object.
3. Geometry can give us about 10 bits of info.
4. Geographical location: 4 bits. Room Name/Outside regions.
5. Height above ground: 1mm to 10,000mm or 10 bits.
6. Distance from nearest wall. another 10 bits.
7. Resting on horizontal surface
8. Color, Blinking lights etc could also provide a few bits.
9. Edges and their relationships.
10. The bot can attach semacode for easy identification. It can also use existing barcodes to gather data.
11. Movement history
12. Characteristic frequencies in texture, movement etc.

Very easily should lead to classification. Once classfied, they must be tracked for their Object Ids. All data collected must be web browsable.

Attributes attached to objects such as cost will determine whether it worth tracking history for it at all or not. A leaf is not worth tracking but a railway ticket is worth tracking.

The whole world would be divided into Home,Compound,Apartment block,Unknown. The Logitude and Latitude will be known through GPS.

It will also store an inventory of objects and track their movement around the house.

On the one hand, a domestic robot does not need to know about insects in the amazon jungles, it does not need to know the capital of mogadishu, even if it gets a doubt in these areas, it can ask a remote system and clarify. If it wont step outside the home, it does not even need to know what a train is or what its parts are.

It needs an extensive knowledge of fruits, vegetables, cleaning processes, friction, viscosity, surface tension, stains, good & bad smells, diseases, dust, strenth of materials, balancing, all books from amazon, It also needs some knowledge of what it can see though windows, insects, spiders, birds, parts of plants and animals, dresses and so on... It should be able to associate anything it physically sees around to its corresponding entry in objects such as bills. It should be able to recognize brands and their corresponding products. It may have to know the nuances of differences b/w 2 models of a cellphone whereas it can avoid to bother about differences b/w lions and tigers. It needs to have an extensive idea about toxicity, rotting, garbage handling etc. Psychology of old persons, anatomy of the body and so on. It needs to be able to understand and co-relate various sounds in the house.

It needs to be able to predict things like what happens when u drop X. Clothes will form a heap while eggs will form a mess.

In search of every #$recipe

2008-03-08T16:15:00.000-08:00

I recently saw a very interesting article about recipies and a symbolic notation for them. by David A Mundie. It was written in 1985. But I think now is the right time for it to be exploited. Most programmers would have learnt programming with comparisons of programs with recipies. We have seen a large number of "cookbooks" for various programming languages!!! Yet we dont have a universal notation for the recipies

I quote from the movie Matrix:

Mouse: Do you know what it really reminds me of? Tasty Wheat. Did you ever eat Tasty Wheat?
Switch: No, but technically, neither did you.
Mouse: That's exactly my point. Exactly. Because you have to wonder: how do the machines know what Tasty Wheat tasted like? Maybe they got it wrong. Maybe what I think Tasty Wheat tasted like actually tasted like oatmeal, or tuna fish. That makes you wonder about a lot of things. You take chicken, for example: maybe they couldn't figure out what to make chicken taste like, which is why chicken tastes like everything.

Thinking of the above quote, I asked myself, "What if every meal I ate tasted different?". How many recipies would I need. 365*100*3 = 100,000 assuming Aubrey de grey is partially right. Assuming that I will like 1/10th of all recipies I will need 1 million recipies. Of course I would not mind a repeat once in a year ;-). So yes, I could have a lot more variety and adventure in my food, if only the person/persons who cook for me can be that flexible. What if McDonalds can sell you a new food item, each time you turn up. How much would it cost? Unless you are into high end stuff, the things that usually cost you are the expertise a lot more than the ingredients, if you wanted such a service from McDonalds how much would it cost if people are replaced by inexpensive robots.

Bootstrapping Artificial Intelligence by Symbiotic co-operation from the Blind Community

2008-03-07T09:13:00.001-08:00

In my previous post, I was talking about the Coffee Making problem Steve Wozniak was using to illustrate the weakness of AI. I just asked myself... A blind person has a fully in tact and perfectly functioning brain. Just the lack of sensory input is putting such a huge hurdle in the quality of life and effectiveness of the blind person... Is it possible to come up with a Human machine combination that can effectively tackle the problem much before a machine is able to independently perform such a task? In the process of doing so we will be stretching the limits of technology. But rather than leaping for the stars all at one, we can learn to fly, then slowly get out of the atmosphere, then land on the moon, then explore the solar syatem and then slowly make our way across the galaxy... We need to have incremental goals that are achievable and at he same time play a role in fulfilling our grand ambitions.

What if you restructure Wozniak's challenge as:

Can a computer with low quality computer vision through a cell phone camera and talking GPS assist a totally blind person to go to a random stranger's house and open the gate and ring the door bell without human/guide dog assistance?

and once that challenge is broken, you can approach the problem:

Can a computer assisted Blind Person go to a stranger's house and make coffee in an unfamiliar environment?

I am thinking of a popular public contest that can be conducted just like the DARPA challenge with live video coverage and all that. To an extent resembling reality TV but a different flavor. Something that hopes to go way beyond entertaining viewers and participants. Something that stretches the boundaries of technologies each time it is played. The prize goes to the guy who wrote the software as well as the blind person who made the coffee.

To the general public many features of AI will take some more time to be interesting enough. If a voice recognition software has an accuracy of 90% it means that it annoys its user once every line is typed. On the contrary if you have systems with that kind of accuracy and give it to blind people they will be more than pleased. If you have a system that will warn you that you are about to use detergent on you hair instead of shampoo, and if that systems gives you one false alarm for every 3 correct ones, a blind person will gladly accept such a system as a blessing from heaven.A blind person can become highly empowered even by a vision system with even a low accuracy. Even if a computer can just report open cupboards, things lying on the floor, cobwebs, ant hills, molds etc growing in undisturbed places and so on, it could be a great help. A few extra false warnings can easily be ignored. The vision systems could report the status of various knobs, blinking lights and so on if the blind person asks the system to read it.

If you are a blind person and you are reading this I would like to know your response to this... If a genie were to appear in front of you and grant you 3 wishes... What would you like a computer to report to you? If a computer could keep reporting more and more things to you, when would you begin to feel that it is talking too much? It is hard for me to imagine this... Do you feel a lot of cognitive overload because most of the data flows in though your ears? But even people with normal vision spend a lot of time on phone and TV and Computer and traffic various other such sources of audio information.

Once the easier challenges are broken, they can be upgraded to more challenging ones like

Can a computer assisted Blind Person go to a stranger's house and cook a randomly chosen recipie?

and once that challenge is broken, you can upgrade to

Can a computer assisted Blind Person go to a stranger's house and babysit a N month old baby for 6 hours?

The intelligence of the person will compensate for mistakes by the system. During the initial phase the blind person gets slowed down. But with improvements I suspect that the situation will significantly improve. If it reports that a bride dressed in a white gown looks like a polar bear, the blind person will tell it not to report any polar bears while in the city as the probability is nearly zero. But if it thinks a rope is a snake, it is better to report it. Better be safe than sorry. CyC knows a LOT. And I bet this would be an amazing way to integrate it with other forms of artificial intelligence like vision and hearing. More details to come soon.

To capture the attention and knowledge of the audience, we can add the facility of webchat and SMS. People can send hints, warnings etc to a central server. The blind person will only hear answers through Text To Speech to questions he/she asks. So one cannot directly guide the person. One can only help the person playing the game by teaching the system about the situation and thus making the system better not only for that particular game but its ability to understand the surroundings in general.

There will be a lot of humor because of the miscommunication and ingenuous ways in which a blind person can exploit the bits and pieces of information provided by the AI system. Depending on the state of the art in such systems, we can adjust the complexity of the task so that it remains challenging and not too easy or not too hard and the system undergoes maximum learning while providing maximum entertainment to the blind participants as well as the audience.

People like to deal with the emotions and feelings of other people in general. Very few care about what a computer knows or not. So this process would attract a fairly large audience as far as I can imagine. And we would be doing genuine work towards the welfare of the blind in addition to continuously improving the state of the art in AI. All knowledge gathered during the show will be opensourced. All Videos, chats etc. will also be opensourced so that further research can be based on this.

Such systems which I expect would cost around $500 to $1000 by 2010 could make a significant contribution to improving the quality of life of blind persons. Over time following Moore's law a computer with the form factor and cost of a PDA with a headmounted camera like the one Justin.tv carries around should be able to assist a blind person as a constant companion. Whether this will be powered by CyC or Not is a tough question. But I suspect it will.

There are already solutions like www.seeingwithsound.com. The hope is to multiply the power of such systems by orders of magnitude.

Just some thoughts... What do you feel? I look forward to your comments!

Bootstrapping AI through crowdsourcing

2008-03-06T00:50:00.000-08:00

We have been hearing a lot about AI. But will it forever be a fantasy like flying cars? Or will we see real stuff happening around us soon? Guys like Ray Kurzweil promise us utopian futures. He argues that, if you just have exponentially increasingly hardware capabilities, the rest will happen by itself. What he says has more sense than my one sentence summary. But when we look at the harsh realities, one is forced to question this optimism.

Here is a response to a question about AI from Steve Wozniak:

Think of the steps that a human being has to do to make a cup of coffee and you have covered basically 10, 20 years of your lifetime just to learn it. So for a computer to do it the same way, it has to go through the same learning, walking to a house using some kind of optical with a vision system, stepping around and opening the door properly, going down the wrong way, going back, finding the kitchen, detecting what might be a coffee machine. You can’t program these things, you have to learn it, and you have to watch how other people make coffee. … This is a kind of logic that the human brain does just to make a cup of coffee. We will never ever have artificial intelligence. Your pet, for example, your pet is smarter than any computer.

Wozniak though he co-invented the macs is not the guy who gets to decide whether AI will happen or not. But his arguments seem sensible and remind you that those who are dreaming of AI could possibly be doing too much wishful thinking. But if someone had predicted the cellphone revolution before inventing the transistor, it would have looked like a joke.

But on the other hand we might not even be aware of stuff that has already happened. Take for example the fact that Cyc which is the brainchild of Doug Lenat has been able to score an A on a chemistry exam. It is still very far away from going to somebody's house and making coffee. But hey even I did not know that till recently ;-). [Will post a reference soon]. If you have heard of the miss carolina fiasco you can say that your pet can sometimes be smarter than even a miss carolina ;-).

But in spite of having such a smart system, if someone asks me, can you think of something that Cyc can do for me today... My mind goes blank. I am not a very bad programmer and have the guts to think beyond the obvious. But still my mind goes blank with the question. On the one hand people have monetized the transitivity of a single simple relationship like a friend of a friend through myspace, facebook etc. But masterpieces like Cyc remain untapped. Can we do something about that?

The Opencyc.org and OpenCyc Foundation were created in the hope of bridging that gap. I used to hang around this place about a year and half ago. Just went back to check out the situation. The hope is that the wisdom of the crowds can be exploited to grow the knowledgebase of Cyc. Initially about 200 man years from extremely advanced academics went into it. Now they plan to do crowdsourcing. Of course Cycorp continues with its original zeal. But they have made it easier for the public to participate at various levels.

But when I ask people around me, they say:

"I am not very confident that ur system will ever demonstrate any "useful" intelligence... I guess it is a nice way for Lenat to extract cash from a stupid government showing it some toy problems. Why do you think I will waste my time and energy on this project to enrich someone else or help try to fulfill this never fulfillable dream".

While the statement can be debated, one should focus on the sentiment. The great hype and subsequent failure of AI to deliver on those lofty goals seems to be the cause. To most people it looks like history just repeating itself.

I quickly realized that people will contribute only if they get the instant gratification. If u contribute to an Open Source project, it will be out and "useful" in months. If you contribute to Wikipedia you are broadcasting your content to the world almost instantaneously. To do the same for AI seems tough. Your content will be out soon. But when will it be "useful?". Projects like Mindpixel illustrate this. There is participation and all that, but it does not get past the critical number to be self sustaining to be useful. Once it crosses the threshold it will thrive even if its founders die. Else it ends up being yet another corpus for PHds to be pulling their hair out. Even if Jimmy Wales dies, Wikipedia will flourish. In fact I suspect improve ;-). Juz kidding...

I am still looking out for ideas relating to such instant gratification by contributing to AI in the short term. I am fully aware that I do more good by playing the factory game than by playing Google Image Labeller. But I still feel the Imagelabeller (ignoring the superior UI) attracts me more. How can an AI game be made compelling enough to make the person feel that "Yay I am doing cool stuff!!!". I am sure it is beyond the UI. In ImageLabeller, I suspect that the labels are indexed and available and improve search results immediately rather than sit in some repository and possibly rust. Most importantly, the problem has the right level of hardness and I keep restlessly asking myself if I am smart enough to label this. In the case of Factory, the questions are usually challenging to a fourth grader or if it ends up being challenging, it is about some absurd thing where I ask, why is this in the repository in the first place? Anyway it has been some time since I played. So things could be better.

On this blog I shall be putting out various ideas and thoughts about how we can achieve tangible improvements in the AI and Semantic Web while simultaneously producing solutions that have immediate benefit.