Not applicable.
Not applicable.
Today, deciphering unknown data has become an increasingly difficult, but necessary, task. As data and storage systems become more complex, programmers are finding that their old matching techniques are inefficient. One common method to identify unknown data is to compare it to known data stored in a database. If the unknown data point matches a point contained in the database, then the unknown data can be identified. However, this process requires one to check the unknown data against every piece of known data in the database. Today, test data is being stored in large multi-dimensional structures that hold enormous quantities. As a result, comparing everything in a database has become quite time consuming and CPU intensive. Furthermore, data itself can be rather large, requiring many complicated computations to match.
For example, this problem arises in the process of audio fingerprinting. Audio fingerprinting is the task of identifying an audio track that is missing or has incorrect metadata. A media player which plays audio files may wish to display the title of the song playing and its artist. Generally, the player will look to a file's metadata in order to determine title and artist. However, such information my be inaccurate or missing from the metadata. In order to determine the needed information, then, the player may try to match a signature or other electronic representation of the song to known signatures or other electronic representations of pre-computed songs. Essentially, the player tries to match the unknown metadata of the song to known metadata, or “fingerprints,” that are stored in a database.
There are inherent performance challenges with matching audio fingerprints, though. For instance, once a song is processed, a 64 floating point number is produced and stored in a large multi-dimensional structure. Each digit of the number may correspond to a different dimension of the database. For audio fingerprinting, it is typical to have a 64 dimension database holding nearly 1,000,000 known fingerprints. Furthermore, finding a match requires comparing the unknown song to every pre-calculated fingerprint in the database. As can be imagined, comparing 1,000,000 64 floating point numbers that are housed in a 64 dimension database is time consuming.
Moreover, the comparison calculation of the unknown song to a known fingerprint is quite complex. It is performed by first mapping the unknown song and known fingerprints into multi-dimensional space. In order to compare songs, each known fingerprint's Euclidean distance from the unknown song is calculated. If the unknown song is close enough to a known fingerprint, it is considered a match. Euclidean distance measurements in multi-dimensional space are quite CPU intensive. Consequently, comparing an unknown song to every fingerprint in a multi-dimensional database is slow.
Therefore, a need arises for a method to accurately locate the section of a multi-dimensional database where a match may exist. If such an area can be pinpointed, only the known data within that area would need to be searched. Finding and searching such an area would greatly reduce the amount of time and processor power needed to identify unknown data. Audio fingerprinting is merely one illustration of the problem at hand. Similar difficulties arise in many applications performing comparisons in multi-dimensional databases. The ability to search such structures more efficiently would greatly reduce the time and CPU usage needed to perform data matching tasks.
This invention provides a method and system for efficiently finding all the data points in multi-dimensional space that are within a given Euclidean distance to an unknown point. Specifically, the invention uses a set of reference points to exclude most of the known data points in a multi-dimensional database that are not within a specific distance to the unknown data point. This in turn eliminates the need to compare many points in the database, thus saving time and processor resources.
First, this invention receives a multi-dimensional database, consisting of pre-computed data mapped in multi-dimensional space. Optimal reference points are then selected, mapped into multi-dimensional space, and used to reduce the number of dimensions of the database. In order to select an optimal set of reference points, the invention calculates the Euclidean distances of each known data point in the database to each reference point in a given set. Those distances are then compared with the unknown point's distances away from each reference point. Points with similar distances away from the reference points as the unknown point are considered possible matches. Accordingly, the optimal set of reference points will return the fewest number of matches. Furthermore, when new data points become available within the database, the reference points can be re-optimized. However, once the optimal reference points are determined, they can be used to match numerous unknown points.
The optimal reference points are used to reduce the number of data points in a database needing to be compared against an unknown data point to identify it. This is done by applying each reference point in order and eliminating data points that are not within a specific distance to the unknown data point. From the remaining data points, the closest one to the unknown point can be considered the unknown's match.
The various embodiments of the present invention will be better understood from the detailed description provided below and from the accompanying drawings of various embodiments of the invention. However, the detailed description and drawings should not be read to limit the invention to the specified embodiments. Rather, these specifics are provided for explanatory purposes to help the invention be better understood.
Embodiments of the present invention provide a system and method for excluding most of the known data points in a multi-dimensional database that are not within a specific distance to an unknown data point. This in turn allows the unknown data point's match to be found much quicker by comparing it to a relatively small quantity of data points.
Specific hardware devices, programming languages, components, processes, and numerous details including operating environments and the like are set forth to provide a thorough understanding of the present invention. In other instances, structures, devices, and processes are shown in block diagram form, rather than in detail, to avoid obscuring embodiments of the present invention. But an ordinary-skilled artisan would understand that embodiments of the present invention may be practiced without these specific details. Further, illustrative data structures used to explain various embodiments of the present invention may be, but are not limited to, databases, spreadsheets, and any other apparatus capable of being a storage medium. Further still, illustrative processors used to explain the various embodiments of the present invention may be, but are not limited to, workstations, personal computers, mainframe computers, or any other suitable processing device. Also, a network may be, but is not limited to either an Internet network, Intranet network, local area network (“LAN”), metropolitan area network (“MAN”), wide area network (“WAN”), or a combination thereof without limiting the scope of the various embodiments of the present invention.
Exemplary Operating Environment
Turning to
The invention is operational with numerous other general purpose or special purpose computing system environments or configurations. Examples of well-known computing system environments and/or configurations that may be suitable for use with the invention include, but are not limited to, personal computers, server computers, handheld or laptop devices, multi-processor systems, microprocessor-base systems, set-top boxes, programmable consumer electronics, network PCs, mini-computers, mainframe computers, distributed computing environments that include any of the above systems or devices and the like.
The invention may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include, but are not limited to, routines, programs, objects, components, and data stores that perform particular tasks or implement particular abstract data types. The invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including memory storage devices.
Turing again to
Communication connections depicted at 26 are exemplary of communication media. Communication media typically embodies computer-readable instructions, data stores, program modules, or any other data in a modulated signal data such as a carrier wave or other transport mechanism and includes any information delivering media. The term “modulated-data signal” means a signal that has one or more of its characteristics set or changed in such a manner so as to encode information in the signal. By way of example, and not limitation, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared, and other wireless media.
Device 10 may also have input device(s) 24, such as keyboard, mouse pen, voice input device, touch input device, etc. Furthermore, device 10 may also have output device(s) 22, such as, but not limited to, display, speakers, printers, etc. All these devices are well known in the art and need not be discussed at length here.
Referring now to
The present invention may be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network 116. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices. In operation, devices 110 communicate data to servers 112 via the network connection.
Dimensional Reduction and Point Matching in Multi-Dimensional Space
As described above, matching unknown data points to known data points in a multi-dimensional data structure is complicated. A simple example of the problem is depicted in
A more efficient method to match unknown data points to known data points in multi-dimensional databases is provided and explained in detail below. An area of the data structure where possible matches exist is located. All data points lying outside of the area are eliminated as possible matches to the unknown point. This allows a system to compare a much smaller number of data points to the unknown, instead of having to compare every point in the database. By targeting and searching fewer data points, processor demands can be reduced.
Any comparison of data to data points in a multi-dimensional database may be addressed. For example, but without limitation, a data point may consist of a series of numbers of any numerical base, as shown in
The flowchart depicted in
Turning to
The second input received is the number of dimensions of the data structure. Eventually, the invention will reduce the dimensions to the number of reference points. Both reference points and dimension reduction will be further discussed below. However, it is important to note that the dimensions are reduced in order to simplify distance calculations in multi-dimensional space. Taking the audio fingerprinting example again, distance calculations for points in the 64 dimension database are computed with respect to the specific reference points used. If six reference points were used, distances would be computed in six dimensions, instead of 64. As can be understood, dimension reduction allows the invention to locate possible matches faster.
The last input given is the number of reference points to use. Again, reference point selection and purpose are discussed at length below. However, the number of reference points to use is an input that is determined by performance and results. Fewer reference points will eventually yield more possible match candidates, thus hindering the invention's ability to eliminate data points. On the other hand, using many reference points will require a greater number of distance calculations. Therefore, the number of reference points to choose is predetermined and given as an input.
After receiving the above inputs, the next step in the pre-computation stage 210 is to select reference points 224. These are points in multi-dimensional space the invention uses to reduce the number of possible matches in the database. Specifically, the reference points triangulate a portion of the database where possible matches to the unknown point exist. All points lying outside the triangulated area are not considered possible matches and are eliminated. Eventually, reference points allow the invention to find a relatively small number of data points that lie a similar distance away from each reference point as the unknown point does. For example, suppose six reference points are used. Eventually, the invention would find every data point in the database that had similar distances (accounting for an error of Δn) to each reference point as the unknown point. This would yield a relatively small number of possible match points, which would significantly reduce the number of data points needing searched.
The positions of the reference points are crucial to effectively eliminate as many possible match candidates as possible, given any unknown point. If the reference points are placed too close together, the benefits of triangulation are lost. If they are too far apart, the triangulation process will yield many results, thus making it ineffective. Therefore, the reference points are optimized in order to maximize their efficiency in reducing data points. Once optimized, the reference points can be used to match numerous unknown data points as they exist. However, if any new data points are added to the database, the reference points can be re-optimized to increase their efficiency.
The order of the reference points is also very important. Because each reference point will eliminate data points in the database as possible match candidates to the unknown, it is ideal for the first reference point to make the largest reduction. After the first reference point reduces a large portion of the data set, subsequent reference points have a much smaller set of data points to concentrate on. In effect, this speeds up queries made to servers and reduces the time needed to locate an unknown point's match. Therefore, it is acceptable to sacrifice the overall efficiency of the reference point set in order to get the first point to make the largest possible reduction.
Moreover, the invention reduces the database's number of dimensions by the reference points. Therefore, instead of calculating distance measurements with respect to every dimension of the data structure, such a query must only be performed with respect to the reference points. Considering the audio fingerprinting example once more, without limitation, if six reference points were used, the 64 dimension database would effectively be reduced to six dimensions. Distance calculations, then, could be done in six dimensions, rather than 64, requiring less processor power and time. As a result, the pre-computation calculations can be performed much quicker using fewer dimensions. These simpler computations require less from a processor.
However, in order to obtain an optimal set, the invention compares results from different groups of reference points. To get a new set of reference points at 318, the invention adds a random multi-dimensional vector to each previous reference point as shown at 320. Then, the distances of each subset point to each new reference point is calculated at 322. Once again, the number of subset points having similar distances away from each reference point as the unknown point are computed, as shown at 324. The number of possible matches resulting from the new reference points is then compared to the number of possible matches given by the previous reference points, at 326 and 328. If the new reference points produce fewer possible match points than the previous reference points, then the new reference points are used, as shown at 328, and the process is repeated. Otherwise, the previous reference points are considered the optimal set. Furthermore, in embodiments the processes of comparing the results of reference points can be parallelized; whereby, the best results achieved by any machine are sent to all other machines.
Once the optimal reference points are determined, they can be used to reduce the number of data points in the database considered possible matches to unknown points. This is the process of stage 212 and is further described in
In the optimization stage, the distance of each known point in the database from each reference point is measured and stored. Then, as part of step 412, the distance of the unknown point to each reference point is determined. All data points having a similar distance to the reference point (as compared to the distance of the unknown data point) are determined. It is considered to be a similar distance if the distances are within Δn of the unknown point. Of the remaining points, the closest one to the unknown is considered its match point 414.
Thus, instead of matching the unknown point against all of the data points in the database, only a small number of possible candidates produced by the optimal reference points need to be compared. Data points having similar distances away from each reference point as the unknown are considered possible match candidates. From these possibilities, a server can determine which one is closest in multi-dimensional space to the unknown point.
In order to visualize exactly how the invention works from start to finish, consider the audio fingerprinting example. Assume a media player receives a .wav file that has missing metadata. As a result, the player cannot determine the song's title and artist. After processing the file, a 64 floating point number is returned that represents the song's identity. To match the song's identity with a known song in the database, the media player accesses, via a network connection, a large database of songs contained in a storage medium. Assume the database is a 64 dimension structure housing 1,000,000 songs that are also represented as 64 floating point numbers. Accordingly, the unknown data point is mapped into 64 dimensional space along with the 1,000,000 known songs.
Instead of comparing the numbers of every song to the number of the unknown song, a smaller search is conducted. Because the invention is asked to match .wav files, it receives a specific Δn and number of reference points to use. Suppose that six reference points are intended to be used. Six points are then mapped alongside the unknown and known data points. Further, the invention randomly selects a subset of the database, consisting of 20,000 data points, for reference point optimization. Within the subset, every data point's distance to each reference point is calculated. Then, the reference points are used in order to reduce data. Using each subset point's calculated distance to each reference point, the invention computes the number of data points having similar distances away from each reference point as the unknown data point. To accommodate for error, the absolute value of the data point's Euclidean distance to the reference point is checked to ensure it is within the error distance Δn. Consider that after all six reference points have reduced the data set, 250 possible match points are left. A small distance is then added to each initial reference point, and the aforementioned process is repeated. This time, assume that three possible match candidates are returned. Since the second set of reference points returned fewer possible matches in the subset than the first set of reference points, the second set is used as the optimal reference points.
Next, the invention measures the distance of the unknown point and every known data point to each reference point. The reference points are applied in order, and data points having distances away from any reference point greater than the unknown point's distance plus Δn are eliminated. This process will return a relatively small number of data points that can be considered possible matches. If one data point remains, it is considered the unknown's match. If multiple points are left, the one closest to the unknown is considered its match.
In sum, this invention significantly reduces the number of pre-calculated data points in a multi-dimensional database needing to be compared against an unknown point in order to find its match. Reference points allow the invention to effectively target and locate a much smaller quantity of points to consider as possible matches. This eliminates the need to compare the unknown point with every possible data point, thus cutting down time and processor requirements to identify an unknown piece of data.
The foregoing descriptions of the invention are illustrative, and modifications in configuration and implementation will occur to persons skilled in the art. For instance, while the present invention has generally been described with relation to
Number | Name | Date | Kind |
---|---|---|---|
6134541 | Castelli et al. | Oct 2000 | A |
6453252 | Laroche | Sep 2002 | B1 |
Number | Date | Country | |
---|---|---|---|
20070011181 A1 | Jan 2007 | US |