31.01.2014 Views

Unsupervised Recursive Sequence Processing - Institute of ...

Unsupervised Recursive Sequence Processing - Institute of ...

Unsupervised Recursive Sequence Processing - Institute of ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

2 <strong>Unsupervised</strong> processing <strong>of</strong> sequences<br />

Let input sequences be denoted by s = (s 1 , . . . , s t ) with entries s i in an alphabet<br />

Σ which is embedded in a real-vector space R n . The element s 1 denotes the most<br />

recent entry <strong>of</strong> the sequence and t is the sequence length.The set <strong>of</strong> sequences <strong>of</strong><br />

arbitrary length over symbols Σ is Σ ∗ , and Σ l is the space <strong>of</strong> sequences <strong>of</strong> length l.<br />

Popular recursive sequence processing models are the temporal Kohonen map,<br />

recurrent SOM, recursive SOM, and SOM for structured data [8,11,39,41]. The<br />

SOMSD has originally been proposed for the more general case <strong>of</strong> tree structure<br />

processing. Here, only sequences, i.e. trees with a node fan-out <strong>of</strong> 1 are considered.<br />

As for standard SOM, a recursive neural map is given by a set <strong>of</strong> neurons n 1 , . . . ,<br />

n N . The neurons are arranged on a grid, <strong>of</strong>ten a two-dimensional regular lattice.<br />

All neurons are equipped with weights w i ∈ R n .<br />

Two important ingredients have to be defined to specify self-organizing network<br />

models: the data metric and the network update. The metric addressed the question,<br />

how an appropriate distance can be defined to measure the similarity <strong>of</strong> possibly<br />

sequential input signals to map units. For this purpose, the sequence entries<br />

are compared with the weight parameters stored at the neuron. The set <strong>of</strong> input signals<br />

for which a given neuron i is closest, is called the receptive field <strong>of</strong> neuron i,<br />

and neuron i is the winner and representative for all these signals within its receptive<br />

field. In the following, we will recall the distance computation for the standard<br />

SOM and also review several ways found in the literature to compute the distance<br />

<strong>of</strong> a neuron from a sequential input. Apart from the metric, the update procedure or<br />

learning rule for neurons to adapt to the input is essential. Commonly, Hebbian or<br />

competitive 1 learning takes place, referring to the following scheme: the parameters<br />

<strong>of</strong> the winner and its neighborhood within a given lattice structure are adapted<br />

such that their response to the current signal is increased. Thereby, neighborhood<br />

cooperation ensures a topologically faithful mapping.<br />

Standard SOM relies on a simple winner-takes-all scheme and does not account<br />

for the temporal structure <strong>of</strong> inputs. For a stimulus s i ∈ R n the neuron n j responds,<br />

for which the squared distance<br />

d SOM (s i , w j ) = ‖s i − w j ‖ 2 ,<br />

s i ∈ R n<br />

is minimum, where ‖ · ‖ is the standard Euclidean metric. Training starts with<br />

randomly initialized weights w i and adapts the parameters iteratively as follows:<br />

denote by n 0 the index <strong>of</strong> the winning neuron for the input signal s i . Assume a<br />

function nhd(n j , n k ) which indicates the degree <strong>of</strong> neighborhood <strong>of</strong> neuron j and<br />

k within the chosen lattice structure is fixed. Adaptation <strong>of</strong> all weights w j takes<br />

1 We will use these two terms interchangeably in the following.<br />

4

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!