Friday, June 6, 2014

Interesting IT patents expiring in 2014

Most patents expire in 20 years after initial filing, leaving the disclosed technology available for everyone. Although most patents filed in 1994 are a bit outdated, there are inventions that still deserve our attention event after 20 years. Here I give you my top of expiring patents.

1. Electronic document verification system and method

This patent covers all digitally signing systems that embed secure information in your document (e.g. using OLE). The embedded information can contain digital signature and/or signator's digitized graphical signature or other graphic image. Here is the example (from the patent):

2. Probabilistic information retrieval networks

Bayesian inference is used to search and rank documents. Consider Bayes theorem:
where p(di|rj) is the probability of selecting document di when a search query contains representation rj. "Representations" can be simple words or more complex structures like phrases, concepts, synonyms, proximities and thesaurus classes. If representations were simple words, there wouldn't be a problem: we can estimate p(rj) by the word frequency, which we can compute during document indexing. For more complex representations this is however not possible: you cannot count every possible phrase, synonym and proximity during indexing because there are way too many combinations. Instead, the authors propose estimating p(rj) in during the document search, when all rj of interest are extracted from the search query. To do this, we compute rj frequencies on random samples of documents. Many other problems are solved as well.

3. Automated, non-invasive iris recognition system and method

Remember the movie called  Minority Report with Tom Cruise? There were cameras everywhere identifying people by their eyes, so that the poor guy had to replace his eyes to remain undetected in the streets. Well, this is it. The person identification system based on eye scanning.
The device features two cameras: one coarse-grained camera for detecting your head and eyes positions, while another fine-grained camera focuses right into your eye with the help of moving mirror.
The iris recognition algorithm is probably outdated, because a lot has happened during the last 20 years in the field. However, the patent is still monopolizing the industry, judging by the immense number of referencing patents.

4. System and method for calculating RAID 6 check codes

Although invented 20 years ago, RAID 6 dominates enterprise sector today. Unlike RAID 5, which only allows one disk failure at a time, RAID 6 allows two disks failed at the same time. In order to understand how does it do recovery, lets first recall how recovery works in RAID 5.
N-disk RAID 5 array divides data in N-1 blocks and adds N-th parity block, which is a XOR-ed data blocks. When a disk fails, its data is recovered by XOR-ing other N-1 disks.
RAID 6 recovery is much more complicated. First define the following operations on M-bit integers:
  1. Addition, a+b, which is a bitwise XOR; subtraction is the same operation;
  2. Multiplication, a*b, yielding an M-bit result; this is not a simple multiplication. Instead, it is a tricky operation involving polynomials in Galois field {0,1} and the boolean AND operation;
  3. Inverse a-1, yielding an M-bit result. This is a unique number such that a-1*a = 1 using the multiplication defined previously;
These operations should form a finite field. Now for a N-disk array, divide data in N-2 blocks and add two check blocks P and Q such that:
for some α and data bits di. Note that addition and multiplication here is as defined before. That is, P is a XOR of data bits, just like in RAID 5. There are the following cases of recovery:
1. Data disk j and the Q check disk fail; restore data disk j like in RAID 5:
2. Data disk j and the P check disk fail; restore data disk j via the formula:

3. Data disks i and j fail:


4. The P and Q check disks fail; reconstruct them from the data disks.

5. Speech efficient coding method

How do you compress human speech? You could of cause treat it as an arbitrary sound: divide into time frames, compute spectrum for each frame and quantize that spectrum. This is how sound compression (e.g. MP3) works. But this is not enough if you want to send it over a low-bandwidth channel like GSM. Mobile telephony employs special speech compression methods so that the radio channel can be shared among as many speakers as possible.
Human speech spectrum possesses few properties that can be used to improve compression rate in comparison to, say, an orchestra. First, there is a voice pitch frequency, which is the main tone. Then, there is a vocal tract that works like a linear filter, transforming the voice. In the frequency domain, the output spectrum S can be modelled as a product of the filter's frequency response T and the original voice's spectrum E:

 S = T E

The original voice E is also called the 'excitation signal', because this is the input signal that excites the filter. For periodic voiced sounds (like vowels), E contains the pitch frequency fp and the related harmonics 2fp, 3fp, etc. The pitch frequency usually varies from 54Hz (low male voice) to 400Hz (high female voice). For unvoiced sounds (like /t/ or /s/), the excitation signal is a white noise, and spectrum E contains some random mess. There are also sounds (like /z/) that appear as periodic in low frequencies and noisy in high frequencies. You may say that /z/ is a voiced sound in some parts of the spectrum and unvoiced in other parts.
To summarize, we can divide the spectrum E into bands fp wide and classify each band as either voiced (V) or unvoiced (UV). 

In the figure to the right, 7A shows the output voice spectrum S, 7B is the filter's frequency response T, and 7C is and excitation spectrum E, which represents an ideal vowel consisting of only V bands. See more real-world spectrums here.
After detecting the filter's coefficients e.g. using LPC analysis, we can encode the whole spectrum using (1) the filter's coefficients, (2) the pitch frequency fp (3) the bit vector containing the V/UV classification result for each band, and (3) amplitude value of each band. This model called Multi-Band Excitation vocoder (MBE) and is in fact more compact way to represent a speech spectrum then coding amplitude for every frequency as in MP3.
The problem of MBE is that in the real world (1) the input sound is noisy and (2) the spectrum of a voiced sound is not strictly periodic. Both problems impacts V/UV classification badly. This patent solves the problems using the following trick: let there be two frequencies, e.g. f1=500-700Hz and f2=3300Hz. If the bands to the left from f1 are classified as V, then, under certain circumstances, the bands between f1 and f2 can be automatically classified as V. This both improves V/UV classification and reduces bitrate.

5. Decoder for decoding ECC using Euclid's algorithm

Reed-Solomon error correcting codes are widely used in the industry: CD/DVD/Blu-Ray discs, barcodes, xDSL, data transmission in space, RAID-6 disk arrays. By introducing redundancy to the transmitted data, Reed-Solomon codes make it possible to recover the transmitted message even if some of the symbols of the codeword were distorted during transmission, or the storage media got corrupted.
Given an input message consisting of k m-bit symbols, the encoder outputs n=k+2t symbols of the codeword, where n2m-1. Reed-Solomon code can recover up to t errors in the codeword - that is, if less then t symbols of the codeword are changed (including both message and control symbols), then the original message can still be recovered.
Reed-Solomon (RS) codes work like this: let the codeword be represented as a polygon in Galois Field of size 2m:

See here how one can construct a finite field of size 2m. The first k coefficients are the same as the input message symbols, and next 2t coefficients are chosen so that c(x) has roots α, α2, ..., α2t, where α is primitive element in GF(2m).
Let r(x) be the received (potentially distorted) message. If r(x) is unchainged, all r(αi) should be zeros. Otherwise, r(x) = c(x) + e(x), where e(x) is the error polygon, and r(αi) = e(αi) (because c(x) = 0 in its root x = αi). The values Si=r(αi) are also called syndromes, and are the values of the error polygon e(x) in the roots α, α2, ..., α2t.
The decoding problem is then to find the coefficients of the polygon e(x) given syndromes Si. The problem can be solved by Euclid's algorithm for finding greatest common polygon divider. The algorithm involves dividing polygons is a loop, and its hardware implementation is tricky. This patent describes the circuit for decoding an RS code.

6. Rapid thumbnail image reconstruction of dct compressed image data

The idea behind this patent is pretty straightforward: use DC coefficients of the JPEG's DCT as the thumbnail pixel values, resulting in shrinkage of the image by the factor of 8 (recall that a DC coefficient is an average of a 8x8 pixel block). Funny that this is exactly what the open-source JPEG library does when you ask to scale an image down by the factor of 8. We were using this principle all these years without knowing it was patentized. In fact, the library goes even beyond that: it can rescale the image during inverse DCT by the ratios of 8-to-N, where N is an integer from 1 to 16. And I doubt the library creators were aware of the patent.

7. System for producing a personal ID card

This patent protects an idea of a pocket-sized card that includes both human-recognizable an machine-readable information, and usually includes the person's name and photo. While this may seem obvious today, it was not that obvious twenty years ago. Weird that Id Technologies LLC, the company now owning the patent, does not seem to be producing such cards currently. It specializes mainly on label printing.

8. Branch target and next instruction address calculation in a pipeline processor


9. Normalization method for floating point numbers

Both patents protect the important building blocks every modern processor consists of. Branch prediction has huge effect on performance, while floating point normalization is a part of every FPU. Now that these blocks are not licensed any more, we can expect some liberation of CPU market in the near future.

10. Method and apparatus for compressing system read only memory in a computing system

BIOS and the setup program in your PC are stored in ROM in a compressed form, so that more code can be fit into a small chip on your motherboard. This patent protects the idea of compressed ROM. The ROM itself contains two parts: the compressed data and uncompressed decompression algorithm. When computer starts, the decompression algorithm executes and decompresses the ROM data into the conventional memory, and then executes the decompressed code.

11. Distributed fingerprints for information integrity verification

A document fingerprint algorithm is proposed that can verify a document using a fingerprint decomposed into several peaces which are distributed among several locations. An adversary can change only a few of those peaces. The majority of the peaces is required to verify the document. The algorithm is based on a error-correcting code.

12. Write anywhere file-system layout

Owned by NetApp Inc, this patent discovers a filesystem featuring fast fileset snapshots and transactions.