I imagine there would be some overhead (e.g. each user transmitting a certain user-specific pattern every 1/2 second, and each AP transmitting a certain AP-specific pattern every 1/2 second), to calibrate the location of each user in AP-space. Fundamentally I think all the calculations involved would be fast linear algebra operations that could be done in hardware on the order of microseconds.
This brings to mind a very interesting idea as well. If the changing physical location of the receiver is causing changes in the composite wireless signal the device is receiving, and the base stations are recalculating and refreshing this all the time very quickly it could potentially be a very reliable and accurate GPS as well.
Actually, I don't think so. This isn't finding your physical location, then computing the correct profile for that location; it empirically probes the performance your location is experiencing, then directly uses that performance information. While you could analyze the results and at least take a stab at the physical location, it's quite possible it will work no better than cell tower triangulation.
Even trying to empirically map certain performance profiles to certain spaces may be impractical if the client radios and antennae performances differ enough to throw off the profiles, which if I understand this properly and given our proclivities for making things as cheap as possible, probably means this won't work either.
The white paper is non-technical, but I think this is the gist:
Currently, if you have multiple users and 1 access point (AP), the users split the bandwidth. Multiple APs and multiple users on the same channel results in split bandwidth as well, since the APs operate independently and interfere with each other.
This proposal uses N APs for N users on the same channel, allowing for full bidirectional use of the channel bandwidth by each user. To send data to N users simultaneously, a central server receives the data, and calculates the signal to send to each AP such that the user receives only the clean signal meant for him post interference. This requires precise localization of the user in AP space, presumably done by having the user transmit a certain pattern at the particular frequency, and measuring the result at each of the channels.
For the N users to transmit simultaneously to the N APs, the data center can take each of the incoming signals from the N APs along with the localization of the users in AP space, and apply linear algebra to unmix the signals into a signal from each user.
I imagine this adds some overhead to each channel in order to maintain precise localizations of each user in AP space.
Hopefully this will improve on the algorithms they have in Lion (which seemed to work ever-so slightly better than OpenCV's face detection and work poorly on profile views).
A lot of the comments below are criticizing "irrational investors" that were "duped" or the product as "vaporware."
This is not the case. Color had a very talented team attempting to attack multiple technically challenging problems, that remain unsolved today.
The first is the discovery of your implicit social network, as defined by your virtual and real-world interactions with others. Facebook currently uses this to determine what information is shown in your News Feed and make friend recommendations, but is not using it to its potential. Google Buzz tried to do this directly via your emails and flopped partly since it did not account for the privacy implications. The ability to transform people's natural interactions into strong recommendations of what they should pay attention to and who should meet each other is still an open problem.
The second is the mapping of real-world events (initially defined by the pictures and people) onto the virtual world. There is potentially a lot of value, both to participants and outsiders, to say (1) who came to real world events, (2) how they interacted, and (3) what happened, while properly dealing with the corresponding ethical implications.
For both of these to work, Color needed a viral social product to gain data and users. They failed on product/market side, especially because they did not have enough focus on "what is the experience we want our users to have the first time they launch the application?" The opportunity remains open, for Color to redeem itself, for the big players to improve their products, or for a new startup to come along and show the world how it's done.
For both of these to work, Color needed a viral social product to gain data and users.
I don't disagree with your analysis, but surely one surefire way to gain social network data is to be bought by someone who has the data already. Google has a pretty good idea who a gmail user's social network is, as evidenced by the Google Plus suggestions.
The implementation of these algorithms is relatively straightforward. The challenge is the state of the art in computer vision is not currently at a point where it is possible to reliably detect 10s-100s of object categories in real time on current systems. It is currently possible to build systems that get decent real-time performance on detecting a few categories concurrently, or offline systems that get around 60% accuracy across hundreds of well-defined categories. Thus, (1) faster general purpose hardware, (2) better algorithms, or (3) running the best algorithms in ASICs designed for CV is necessary. The top labs now typically use GPU clusters to train / run their algorithms, with the computationally expensive stages usually being feature extraction and/or classifier training.
Google Predict (http://code.google.com/apis/predict/) offers a general machine learning API geared towards those who want to apply machine learning to their applications without subject-specific knowledge. I've not used it so I can't speak to its accuracy, but it is not geared towards computer vision and I imagine it would fail miserably at such tasks (since computer vision is highly dependent on domain-specific feature extraction techniques), and I imagine it performs well at NLP tasks. The primary limitation of such a system is that it acts as a black box - you throw data in and get answers out without any knowledge of the process behind it.
This black-box model is limiting for three major reasons. First, depending on the domain, incorporating domain-specific knowledge can greatly improve performance. Secondly, it is hard to understand the limitations of such a system. Many ML algorithms can fail catastrophically when the input is substantially different from the training data, and the black box makes it hard to understand when the system is likely to fail and adjust accordingly. Third, in many cases you face a tradeoff involving speed, memory, and classification/regression performance. This tradeoff is automatically determined for you and is not transparent.
I've been considering a general ML system that offers an API similar to Google Predict, yet is transparent in the feature extraction / model selection stages for those that would benefit from digging deeper into the system. Is this something that you would pay for?
Specifically for computer vision, there's a variety of startups and companies working on providing a system for object recognition and classification. One example is http://www.numenta.com/, though when I tried there software about a year ago it did not seem to function very well compared to the state of the art. Others that are making visual search type applications include http://www.tineye.com/ and http://www.kooaba.com
There is no such thing as a "best" data mining algorithm. Almost all the advantages you mentioned for decision trees, a form of recursive binary partitioning, applies to a greater extent to Random Forests, which are bootstrapped decision trees that only consider a subset of features at each node.
Examples of domains where decision trees perform poorly include:
-Low amount of data
-Domains where you have extra knowledge about the data (such as some features coming from certain probability distributions) that you can incorporate into classifiers.
Decision trees work well in a variety of applications, but that does not make them the "best" algorithm, and it is rare that a classical decision tree provides state of the art performance on any given data set. http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.122...
Agree completely, had just copied the quote and was about to paste it to say the same thing. When we consider all the BS patents that have been rubber-stamped by the PTO, this is a very scary proposal.
First paragraph doesn't make since - if there are 5 flips, there are 2^5=32 possible outcomes. If the "odds are low that even one person in the audience guessed it" then I'd expect less than 16 people to be in the audience. However, "about a dozen people" did guess it, implying that the audience is in the hundreds (there are >10 "random" looking sequences of the 32).
