<div dir="ltr"><span style="color:rgb(0,0,0);font-family:-webkit-standard">Dear All:</span><br><div class="gmail_quote"><div dir="ltr"><div dir="ltr"><div class="gmail_quote"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div class="m_-1956686164521569017gmail-m_4867929685797778472gmail-page" title="Page 1"><div class="m_-1956686164521569017gmail-m_4867929685797778472gmail-layoutArea"><div class="m_-1956686164521569017gmail-m_4867929685797778472gmail-column"><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>If we do not get back to much earlier non-computational studies, it was Alan Turing 1936 who proposed what is now called universal Turing machines (Alan Tring's meaning "computable numbers" has been enriched later to all possible computer programs). Alan Turing 1950 asked: "Can machines think?". Let us use 1950 as the starting year for scientists to start this extremely challenging computational question. This is probably of one of the the most important issues for modern time. </span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>John Tsotsos and his coworker has a recent review Jan. 13, 2018, "A Review of 40 Years of Cognitive Architecture Research: Core Cognitive Abilities and Practical Applications," available from ArXiv. From this review, we can see that the problem we face is an extremely challenging one. </span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>Nov. 10, 2018, Richard Loosemore wrote on this post: "Why is the question of how concepts relate to neurons so scandalously incoherent, even after this field has been talking about it for at least 35 years?" Loosemore, R.P.W. & Harley, T.A. (2010) wrote: "If you look through a journal such as Science or Nature, you will find that most articles on psychology contain or refer to imaging. And what now passes as psychology in the popular press is mostly reports on brain imaging studies. The brain is back in cognitive psychology. But is this change for the better?" </span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>Nov. 11, 2018, Asim Roy wrote: "I would argue that localist representation is synonymous with symbolic systems." </span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>All the above questions are good, but such partial questions only lead to partial solutions that do not address more fundamental issues, only delaying them. </span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>I proposed to ask a more holistic question below. This question has been positively answered through (1) a theory, (2) a fully detailed algorithm, (3) an optimality proof, (4) experiments with impressive performances verified by several independent groups during the AIML Contest 2016.</span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>"How machines auto-program for general purposes (through a single lifetime learning in the natural world)?" This question is concise but avoids many pitfalls that many partial questions left to us and misled us.</span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>The following freely available technical report provides a solution to this question: Juyang Weng: "A Model for Auto-Programming for General Purposes" <a href="https://arxiv.org/abs/1810.05764v1" target="_blank">https://arxiv.org/abs/1810.05764v1</a></span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>Of course, this report did not solve all problems. The following are some key points:</span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>(1) Liberate symbols to overcome the limitation that a set of symbols is handcrafted only after a specific task is given. Weng modeled that emergent numeric vectors correspond to muscles neurons on the motor end. These numeric vectors liberate (i.e., without the limitations of using) symbols because they emerge automatically. </span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>(2) As a bridge for understanding, Weng used symbols to explain how a DN learns emergent Universal Turing machines, which is a model of modern general-purpose computers. </span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>(3) Generality: Such emergent vectors on the motor end correspond to all motor-expressible (e.g., spoken or hand-written) spatiotemporal concepts (not just mental states), such as states, actions, plans, goals, intents, costs, plans, goodness, badness, and novelties. </span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>(4) Skull-closed: the model (DN)'s skull-internal learning is always unsupervised, similar to the skull that closes the human brain, off limit to human teachers. </span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>(5) Top-down weights: Each hidden neuron has two types of weight vectors, top-down weight vectors (this is fundamentally new compared with all feedforward and recurrent neural networks) and bottom-up weight vectors (typically local). </span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>(6) The model is without any rigid boarders between “lower processing” and “high processing” where “lower” sensory inputs (e.g., vision, audition, and text) and “higher” context inputs (e.g., cognition, planning, and natural languages) are tightly integrated.</span></font></div><div><font color="#000000" face="-webkit-standard"><span><br></span></font></div><div><font color="#000000" face="-webkit-standard"><span>-John Weng</span></font></div></div></div></div></div></div></div></div></div></div></div></div></div></div></div></div></div></div>
</div></div></div>
</div></div>