Phoneme discrimination using connectionist networks

Objective: The aim of this project was to develop the Arabic CAPT (A-CAPT), a Standard Arabic version of the CHEAR auditory perception test (CAPT) that assesses consonant perception ability in children. Method: This closed-set test was evaluated with normal-hearing children aged 5 to 11 years. Development and validation of the speech materials were accomplished in two experimental phases. Twenty-six children participated in phase I, where the test materials were piloted to ensure that the selected words were age appropriate and that the form of Arabic used was familiar to the children. Sixteen children participated in phase II where test–retest reliability, age effects, and critical differences were measured. A computerized implementation was used to present stimuli and collect responses. Children selected one of four response options displayed on a screen for each trial. Results: Two lists of 32 words were developed with two levels of difficulty, easy and hard. Assessment of test–retest reliability for the final version of the lists showed a strong agreement. A within-subject ANOVA showed no significant difference between test and retest sessions. Performance improved with increasing age. Critical difference values were similar to the British English version of the CAPT. Conclusions: The A-CAPT is an appropriate speech perception test for assessing Arabic-speaking children as young as 5 years old. This test can reliably assess consonant perception ability and monitor changes over time or after an intervention.

Download Full-text

Using Noise to Compute Error Surfaces in Connectionist Networks: A Novel Means of Reducing Catastrophic Forgetting

Neural Computation ◽

10.1162/08997660260028700 ◽

2002 ◽

Vol 14 (7) ◽

pp. 1755-1769 ◽

Cited By ~ 6

Author(s):

Robert M. French ◽

Nick Chater

Keyword(s):

Feedforward Networks ◽

First Derivative ◽

Error Surface ◽

Connectionist Networks ◽

New Information ◽

Catastrophic Interference ◽

Derivative Information

In error-driven distributed feedforward networks, new information typically interferes, sometimes severely, with previously learned information. We show how noise can be used to approximate the error surface of previously learned information. By combining this approximated error surface with the error surface associated with the new information to be learned, the network's retention of previously learned items can be improved and catastrophic interference significantly reduced. Further, we show that the noise-generated error surface is produced using only first-derivative information and without recourse to any explicit error information.

Download Full-text