Re: Two challenges related to KR of scientific papers and books

> On 21 Nov 2022, at 09:36, Dave Raggett <dsr@w3.org> wrote:
> 
> Anyone interested in this should take a look at the capabilities of Wolfram Mathematica, which is a mature closed source system for working with math. Some open source alternatives include SageMath, Gnu Octave, Jupyter and Maxima, e.g.
> 
>> Maxima is a system for the manipulation of symbolic and numerical expressions, including differentiation, integration, Taylor series, Laplace transforms, ordinary differential equations, systems of linear equations, polynomials, sets, lists, vectors, matrices and tensors. 
> 
> 
> An AGI agent would go a lot further, and it would be interesting to consider how to train an AGI to give it good mathematical skills, and enable it to improve further by studying the mathematical literature.  


It is unlikely that such an AGI can be realised using handcrafted knowledge representations, sorry, Paola!

I am therefore looking into how to architect reasoners that operate on latent semantics as an enabling technology. Deep learning has proven effective in developing large language models, but lack reasoning skills, e.g. take a look at the following example produced using BLOOM, where the human supplied text is in bold, and the text generated by the large language model is in italics.

There is one blue ball, two green balls and five red balls. How many balls are there in total? The first possible answer is '7' because there is one blue ball and six other balls. The second possible answer is '4', because the first ball is blue, two of the other six balls are green and the remaining four balls are red.

Which sounds like a student trying to cover up their lack of mathematical competence!

In principle, it should be straightforward to train a network to perform operations over latent semantics for natural language and for images, assuming we can devise an appropriate network architecture. The training dataset could be autogenerated.  Mathematics is probably an easier domain for this than software engineering, but who knows!

Dave Raggett <dsr@w3.org>