Nick Byrd, Ph.D.<p>Why <a href="https://nerdculture.de/tags/DeepNeuralNetworks" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>DeepNeuralNetworks</span></a> need <a href="https://nerdculture.de/tags/Logic" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>Logic</span></a>:</p><p>Nick Shea (<a href="https://nerdculture.de/tags/UCL" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>UCL</span></a>/#Oxford) suggests</p><p>(1) Generating novel stuff (e.g., <a href="https://nerdculture.de/tags/Dalle" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>Dalle</span></a>'s art, <a href="https://nerdculture.de/tags/GPT" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>GPT</span></a>'s writing) is cool, but slow and inconsistent.</p><p>(2) Just a handful of logical inferences can be used *across* loads of situations (e.g., <a href="https://nerdculture.de/tags/modusPonens" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>modusPonens</span></a> works the same way every time).</p><p>So (3) by <a href="https://nerdculture.de/tags/learning" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>learning</span></a> Logic, <a href="https://nerdculture.de/tags/DNNs" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>DNNs</span></a> would be able to recycle a few logical moves on a MASSIVE number of problems (rather than generate a novel solution from scratch for each one).</p><p><a href="https://nerdculture.de/tags/CompSci" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>CompSci</span></a> <a href="https://nerdculture.de/tags/AI" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>AI</span></a></p>