Agreement measures are useful metrics to both compare different evaluations of the same diagnostic outcomes and validate new rating systems or devices. While many of them have been proposed in the literature so far, Cohen's n is still the de facto standard in gauging the agreement.Information Agreement (IA) is a novel two-observers information-theoretic-based metric introduced to overcome all the limitations and alleged pitfalls of Cohen's n. It offers an operative meaning to the agreement since it measures - in both dichotomous and multi-value ordered-categorical cases-the information shared between two raters through the virtual diagnostic channel connecting them: the more information exchanged between the raters, the higher their agreement. Unfortunately, this measure is only able to deal with agreement matrices whose values are all strictly positive numbers.This work extends IA by admitting also 0 as a possible value for the entries of an agreement matrix. Moreover, a Python software library to compute the extended version of IA, together with some of the most used agreement measures, is presented and tested.

One application of the P300 brain electric signal is sentence spelling, which enables subjects who have lost control of their motor pathways to communicate by selecting characters in a matrix containing all alphabet symbols. This technology still suffers from both low communication/high error rates. A P300 speller, named PolyMorph, which jointly introduces the selection matrix polymorphism (reducing the matrix size by removing useless symbols) and sentence-based predictions (which forecast words on the basis of language statistics) is presented. This is accomplished by using a custom dynamic knowledge-base, tailored to the subject lexicon, and updated in real time with the selections of the subject. The effectiveness of the presented speller is measured in vivo and in silico. The results suggest that the use of PolyMorph increases the number of spelt characters per time unit and reduces the error rate.

We study the composition of messages in an encoding tree for a Tunstall code, and, more generally, in a tree whose skewness is bounded. For such trees a sort of "law of large numbers" holds true; actually, we provide a direct and converse coding theorem for variable-length to block length source codes, when a vanishing error probability is allowed.

In the first part of this paper, we tackle the case where a variable length-to-block Tunstall code is used to encode the wrong source (miscoding). It turns out that, exactly as happens in the case with Huffman coding, the asymptotic excess rate is given by the informational divergence between the probability distribution ruling the source and the probability distribution for which the Tunstall code had been devised. We also prove asymptotic equality between the individual rate of a codeword and the corresponding self-information. This allows us to bound the maximal and the minimal length in a Tunstall tree. In the second part of the paper we study the case where the probability distribution of the source changes in time. If this happens, it is necessary to frequently update the current code, to ensure that the optimality conditions concerning its rate are met. This coding procedure is known as adaptive coding. We propose some schemes for adaptive Tunstall coding, based on the structure of the coding tree, on the "ordering property", analogous to Gallager's (1978) sibling property of the Huffman code, and on the "Tunstall regions", analogous to the attraction regions due to Longo and Galasso (1982).

An application of the neural Hopfield model to the source (optimal) coding problem is presented. This application gives the structure of a Hopfield network, capable of calculating the lengths of the codewords that must be associated to a source alphabet so as to minimize their average length. The case of a stationary and memoryless block-to-variable-length source coding is discussed.>

Some results on variable-length-to-variable-length source coding are presented. A sufficient criterion for the asymptotic optimality of the generic V-V code is given. This also allows the study of the Tunstall-Huffman scheme performance. Then, a nonoptimal greedy approach to prefix codes is described, based on the informational divergence pseudometric and on the Gallager algorithm minimization of the current rate, associated with the subsequent extensions of the source probability distribution. Although optimality is not always reached, this technique can be usefully employed to improve the Tunstall-Huffman concatenation.>