Databases Reference
In-Depth Information
Less attention has been paid to the aspect of neuron connectivity in the net-
work. On this perspective, we can see a number of significant improvements
in the accuracy of conventional pattern recognition schemes, such as Hop-
field network. Nevertheless, the synaptic plasticity effect of tightly coupled
learning algorithms, such as Hebbian learning, limits the scalability of these
schemes. In relation to pattern recognition, to move beyond the boundaries of
Internet-scale environments, it is important for us to revisit the fundamentals
of neural networks. Throughout the discussions presented in this topic, sev-
eral important concepts of in-network processing for pattern recognition were
presented. These included one-shot learning (Chapter 3), the hierarchical pro-
cessing model (Chapter 4), and the divide-and-distribute approach (Chapter
5). These concepts build up the foundations of a DPR approach for Internet-
scale recognition.
With the ability to expand recognition beyond the sequential train-
validation-test approach of existing pattern recognition schemes, DPR enables
recognition processes to be performed in a parallel and distributed manner.
The distributed multi-feature recognition approach, discussed in Chapter 7,
provides a scalable means of implementing recognition procedures on multi-
ple data features. With the expanding network resource availability, one is
capable of implementing pattern recognition using an unlimited number of
features. This factor is important, as we can see from the McGurk effect [103],
because certain features can be retrieved or detected only by using a combi-
nation of features. Instead of identifying information, we are able to generate
information using distributed information processing.
10.3.2 Recognition as Commodity
Apart from changing the perspective of recognition, this topic was also
intended to deliver the concept of pattern recognition as a commodity for
information processing applications. As was demonstrated in Chapters 8 and
9, DPR schemes, such as the DHGN, can be deployed in different types of
computational networks. This adaptive feature for network granularity en-
ables recognition processes to be treated as a generic service or commodity
for different types of applications. We have demonstrated the use of DPR
in face recognition and distributed event detection. The unique approach of
the recognition process via in-network computations allows recognition to be
performed regardless of the types of data being used. Thus, this approach
enhances the scalability of the pattern recognition approach by taking into
account the structure and resources available in a particular network.
With this concept of recognition as a commodity, we are able to fully uti-
lize current and future technology, such as cloud computing, which has been
developed with a service-oriented architecture (SOA). We can conceptualize
pattern recognition as a cloud service that can be deployed in different types of
analytical and information processing applications ranging from a simple im-
