Big Data and the SP Theory of
Intelligence
Varsha PrabhakaranS8 CSE B
Roll No: 43
1
Contents
Introduction
SP Theory of Intelligence
Problems of Big Data
Volume
Efficiency
Transmission
Variety
Veracity
Visualization
2
Introduction
SP theory of intelligence be applied to the management
and analysis of big data
Overcomes the problem of variety in big data.
Analysis of streaming data- velocity
Economies in the transmission of data
Veracity in big data.
Visualization of knowledge structures and inferential
processes3
SP Theory of Intelligence
4
SP Theory of Intelligence Designed to simplify and integrate concepts across artificial intelligence, mainstream computing, and human perception and cognition.
Product of an extensive program of development and testing via the SP computer model.
Knowledge represented with arrays of atomic symbols in one
or two dimensions called “patterns”.
Processing are done by compressing information
Via the matching and unification of patterns.
Via the building of multiple alignments .
5
Benefits of the SP Theory
Conceptual simplicity combined with descriptive and
explanatory power across several aspects of intelligence.
Simplification of computing systems, including software.
Deeper insights and better solutions in several areas of
application.
Seamless integration of structures and functions within and
between different areas of application
6
SIMPLIFICATION OF COMPUTING SYSTEMS
7
MULTIPLE ALIGNMENT: A CONCEPTBORROWED FROM BIOINFORMATICS
8
Multiple Alignment
The system aims to find multiple alignments that enable a
New pattern to be encoded economically in terms of one or
more Old patterns
Multiple alignment provides the key to:
Versatility in representing different kinds of knowledge.
Versatility in different kinds of processing in AI and mainstream
computing.
9
Multiple Alignment
S → NP V NP
NP → D N
D → t h i s
D → t h a t
N → g i r l
N → b o y
V → l o v e s
V → h a t e s
10
Multiple Alignment
11
Multiple Alignment
12
S 0 1 0 1 0 #S
Multiple Alignment
Compression difference:
CD = BN-BEBN :total number of bits in those symbol in the New pattern that are aligned with Old symbols in the alignment
BE :the total number of bits in the symbols in the code pattern
Compression ratio:
CR = BN/BE;
13
Multiple Alignment
BN is calculated as: h
BN = Ʃ Ci i=1
Ci is the size of the code for ith symbol in a sequence, H1...Hh, com- prising those symbols within the New pattern that are aligned with Old symbols
14
Multiple Alignment
BE is calculated as:
s
BE = Ʃ Ci
i=1
where Ci is the size of the code for ith symbol in the sequence of s symbols in the code pattern derived from the multiple alignment.
15
Multiple Alignment
16
17
18
Big Data
19
Problems of Big Data and Solutions
Volume: big data is … BIG!
Efficiency in computation and the use of energy.
Unsupervised learning: discovering ‘natural’ structures in data.
Transmission of information and the use of energy.
Variety: in kinds of data, formats, and modes of processing.
Veracity: errors and uncertainties in data.
Interpretation of data: pattern recognition, reasoning
Velocity: analysis of streaming data.
Visualization: representing structures and processes
20
Volume: Making Big Data Smaller
“Very-large-scale data sets introduce many data management
challenges.”
Information compression.
Direct benefits in storage, management and transmission.
Indirect benefits
efficiency in computation and the use of energy
unsupervised learning
additional economies in transmission and the use of energy
assistance in the management of errors and uncertainties in data
processes of interpretation.
27
Energy, Speed and Bulk
In the SP theory, a process of searching for matching patterns
is central in all kinds of ‘processing’ or ‘computing’.
This means that anything that increases the efficiency of
searching will increase computational efficiency and,
probably, cut the use of energy:
Reducing the volume of big data.
Exploiting ***probabilities***.
Cutting out some searching. 22
Efficiency via Reduction in Volume
Information compression is central in how the SP system works:
Reducing the size of big data.
Reducing the size of search terms.
Both these things can increase the efficiency of searching, meaning gains in computational efficiency and cuts in the use of energy.
23
Efficiency Via Probabilities
24
Efficiency Via Probabilities
25
26
Efficiency Via Probabilities
26
27
Efficiency Via Probabilities
27
Efficiency Via Probabilities
Statistical knowledge flows directly from:
Information compression in the SP system and
The intimate connection between information compression and
concepts of prediction and probability.
There is great potential to cut out unnecessary searching, with
consequent gains in efficiency.
Potential for savings at all levels and in all parts of the system
and on many fronts in its stored knowledge.
28
Efficiency via a Synergy with Data-Centric Computing
29
Efficiency via a Synergy with Data-Centric Computing
In SP-neural, SP patterns may be realized as
neuronal pattern assemblies.
There would be close integration of data and
processing, as in data-centric computing.
Direct connections may cut out some
searching30
Unsupervised learningLossless compression of a body of information
Information compression, or “minimum length encoding”
remains the key.
Matching and unification of patterns
SP computer model has already demonstrated an ability to
discover generative grammars, including segmental
structures, classes of structure, and abstract patterns.
For body of information, I, the products of learning are:
a grammar (G) and an encoding (E) of I in terms of G31
Product of Learning
32
Transmission of Data
• By making big data smaller (“Volume”).
• By separating grammar (G) from encoding (E), as in some
dictionary techniques and analysis/synthesis schemes.
• Efficiency in transmission can mean cuts in the use of energy.
33
Transmission of Data
34
Transmission of Data
Simplicity of a focus on the matching and unification of
patterns.
Aims to discover structures that are, quotes, “natural”.
Brain-inspired “DONSVIC” principle can mean relatively
high levels of information compression.
Potential for G to include structures not recognized by most
compression algorithms, such as: Generic 3D models of objects and scenes.
Generic sequential redundancies across sequences of frames. 35
Overcoming Problems of Variety of Big Data
Diverse kinds of data: the world’s many languages, spoken or written; static and moving images; music as sound and music in its written form; numbers and mathematical notations; tables; charts; graphs; networks; trees; grammars; computer programs; and more.
There are often several different computer formats for each kind of data. With images, for example: JPEG, TIFF, WMF, BMP, GIF, EPS, PDF, PNG, PBM, and more.
Adding to the complexity is that each kind of data and each format normally requires its own special mode of processing.
THIS IS A MESS! It needs cleaning up.
Although some kinds of diversity are useful, there is a case for developing a universal framework for the representation and processing of diverse kinds of knowledge (UFK).
36
Universal Framework for the Representation and Processing of Knowledge(UFK)
Potential benefits of a UFK in: ● Learning structure in data ● Interpretation of data; ● Data fusion; ● Understanding and translation of natural languages; ● The semantic web and internet of things; ● Long-term preservation of data; ● Seamless integration in the representation and processing of diverse kinds of knowledge.
Most concepts are an amalgam of diverse kinds of knowledge (which implies some uniformity in the representation and processing of diverse kinds of knowledge).
The SP system is a good candidate for the role of UFK because of its versatility in the representation and processing of diverse kinds of knowledge.
37
How Variety Hinders LearningDiscovering the association between lightning and thunder is likely to be difficult when: Lightning appears in big data as a static image in one of several formats; or in a moving image in one of several formats; or it is described, in spoken or written form, as any one of such things as “firebolt”, “fulmination”, “la foudre”, “der Blitz”, “lluched”, “a big flash in the sky”, or indeed “lightning”.
Thunder is represented in one of several different audio formats; or it is described, in spoken or written form, as “thunder”, “gök gürültüsü”, “le tonnerre”, “a great rumble”, and so on.
If learning and discovery processes are going to work effectively, we need to get behind these surface forms and focus on the underlying meanings. This can be done using a UFK.
38
Veracity
“In building a statistical model from any data source, one must often deal with the fact that data are imperfect. Real-world data are corrupted with noise. … Measurement processes are inherently noisy, data can be recorded with error, and parts of the data may be missing.”
In tasks such as parsing or pattern recognition, the SP system is robust in the face of errors of omission, addition, or substitution.
39
Veracity
40
Veracity
When we learn a first language (L):
We learn from a finite sample.
We generalize (to L) without over-generalising.
We learn ‘correct’ knowledge despite ‘dirty data’.
41
Veracity
For any body of data, I, principles of minimum-length encoding
provide the key:
Aim to minimize the overall size of G and E.
G is a distillation or ‘essence’ of I, that excludes most ‘errors’
and generalizes beyond I.
E + G is a lossless compression of I including typos etc but
without generalizations.
Systematic distortions remain a problem.42
Interpretation of Data
Processing I in conjunction with a pre-established grammar (G) to create a
relatively compact encoding (E) of I
Depending on the nature of I and G, the process of interpretation may be
seen to achieve:
Pattern recognition
Information retrieval
Parsing and production of natural language
Translation from one representation to another
Planning
Problem solving43
Velocity: Analysis of Streaming Data
In the context of big data, “velocity” means the analysis
of streaming data as it is received.
“This is the way humans process information.”
This style of analysis is at the heart of how the SP
system has been designed.
Unsupervised learning.
44
Visualizations
The SP system is well suited to visualization for these reasons:
Transparency in the representation of knowledge.
Transparency in processing.
The system is designed to discover ‘natural’ structures in data.
There is clear potential to integrate visualization with the
statistical techniques that lie at the heart of how the SP system
works.
45
Conclusion
Designed to simplify and integrate concepts across artificial intelligence, mainstream computing, and human perception and cognition, has potential in the management and analysis of big data.
The SP system has potential as a universal framework for the representation and processing of diverse kinds of knowledge (UFK), helping to reduce the problem of variety in big data
the great diversity of formalisms and formats for knowledge, and how they are processed.
46
Bibliography
www.cognitionresearch.org/sp.htm .
Article: “Big data and the SP theory of
intelligence”, J G Wolff, IEEE Access, 2, 301-
315, 2014.
47
48