[Link] Word-vector based DL system achieves human parity in verbal IQ tests

A re­search team in China has cre­ated a sys­tem for an­swer­ing ver­bal anal­ogy ques­tions of the type found on the GRE and IQ tests that scores a lit­tle above the av­er­age hu­man score, per­haps cor­re­spond­ing to an IQ of around 105 or so. This im­proves sub­stan­tially on the re­ported SOTA in AI for these types of prob­lems.

This work builds on deep word-vec­tor em­bed­dings which have led to large gains in trans­la­tion and many NLP tasks. One of their key im­prove­ments in­volves learn­ing mul­ti­ple vec­tors per word, where the num­ber of spe­cific word mean­ings is sim­ply grabbed from a dic­tio­nary. This is im­por­tant be­cause ver­bal anal­ogy ques­tions of­ten use more rare word mean­ings. They also em­ploy mod­ules spe­cial­ized for the differ­ent types of ques­tions.

I vaguely re­mem­ber read­ing that AI sys­tems already are fairly strong at solv­ing vi­sual raven-ma­trix style IQ ques­tions, al­though I haven’t looked into that in de­tail.

The multi-vec­tor tech­nique is prob­a­bly the most im­por­tant take away for fu­ture work.

Even if sub­se­quent fol­low up work reaches su­per­hu­man ver­bal IQ in a few years, this of course doesn’t im­me­di­ately im­ply AGI. Th­ese types of IQ tests mea­sure spe­cific abil­ities which are cor­re­lated with gen­eral in­tel­li­gence in hu­mans, but these spe­cific abil­ities are only a small sub­set of the sys­tems/​abil­ities re­quired for gen­eral in­tel­li­gence, and prob­a­bly rely on a small­ish sub­set of the brain’s cir­cuitry.