Generalization at Retrieval Using Associative Networks with Transient Weight Changes

Abstract Without having seen a bigram like “her buffalo”, you can easily tell that it is congruent because “buffalo” be aligned with more common nouns “cat” or “dog” have been in contexts cat” dog”—the novel structurally aligns representations memory. We present new class of associative nets we call Dynamic-Eigen-Nets , and provide simulations show how they generalize to patterns are the training domain. Linear-Associative-Nets respond same pattern regardless input, motivating introduction saturation facilitate other response states. However, models using cannot readily novel, but patterns. address this problem by dynamically biasing eigenspectrum towards external input temporary weight changes. demonstrate two-slot Dynamic-Eigen-Net trained on text corpus provides an account judgment-of-grammaticality lexical decision tasks, showing better capture syntactic regularities from compared Brain-State-in-a-Box Linear-Associative-Net. end simulation sensitive violations introduced bigrams, even after associations encode those bigrams deleted Over all simulations, reliably outperforms propose as at retrieval, instead encoding, through recurrent feedback.