Home > Uncategorized > Weekly QuEST Discussion Topics and News, 18 Sept

Weekly QuEST Discussion Topics and News, 18 Sept

QuEST for 18 Sept 2015:

We will have a discussion on thinking – specifically starting with a presentation by Geof Hinton on:

Aetherial Symbols

Geoffrey Hinton

University of Toronto

&

Google Inc.

Kosslyn thought that we process images by operating on

pixels in our head. Pylyshyn thought that we process

sentences by operating on symbols in our head.

  • They were both making the same naive mistake.

– What we have in our heads is not a cleaned up

version of the input. ** this begs the question why we are seeking to populate an internal representation for our computer agents with cleaned up versions of the  sensory data ** There are no pixels or symbol

strings in the head.

  • All we have in our heads is big activity vectors that cause

more big activity vectors.

– So why did people think we have symbol strings or

images in the head?

– Maybe they misunderstood how everyday language

refers to internal states.

How does intuitive reasoning work?

  • How can a neural net infer the answer to a

simple question without doing any explicit,

sequential reasoning?

  • Represent each word by a large vector of

features each of which has causal effects.

– The combined effects of all the features

capture knowledge that symbolic AI would put

into an explicit rule of inference.

Natural reasoning

  • Once we can turn sentences into thought

vectors, we can learn to predict a thought vector

from previous thought vectors.

– This should allow us to model natural

reasoning, but it hasn’t yet been done.

—- this last point is where we want to focus –

The variables used to create/use these thought vectors:

  • Nearly all artificial neural nets make do with just two

kinds of variables:

– Neural activities are used to represent what the net is

currently thinking about.

– Weights are used to store the long-term knowledge.

Will we require a new type of variable to do QuEST thought vectors?

Let’s architect (in our discussions during the QuEST meeting) a dual process theory compliant model using a deep learning CNN / RNN infrastructure.  In the DPT solution the units of cognition in the 2nd system, vocabulary manipulated and used to populate working memory are what we call qualia.  Qualia are the units of cognition used in working memory.  Any agent that implements a DPT solution and generates a DPT sys2 where the units of cognition (vocabulary of working memory) are compliant with our QuEST tenets of Structural coherence, Situated and simulated are called QuEST agents.

As a means to prime the discussion imagine a CNN/RNN system for generating narrative descriptions of images/video snippets.  As currently implemented in the AC3 system.  The thought vectors (weight space representation of the sentences) are generated in a reflexive manner.  If you will using the terminology of Evans / Stanovich the autonomous mind.  Those thought vectors can be used to generate an output (reflexive response) for the agent but can also be used to create a working memory representation.  This is where our discussion will focus.

news summary (25)

Advertisements
Categories: Uncategorized
  1. No comments yet.
  1. No trackbacks yet.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: