I am really disappointed that my AGI 2019 paper has been rejected. The
reasons given by the reviewers are very superficial and vacuous, and given
that I have posted my presentation slides here which explained the theory
in very simple terms, and they have not given me a chance to explain any
unclear details to them (it could be argued that they lack certain basic
AGI background notions, and it's not my fault to omit them). Either the
reviewers don't understand my ideas or they are biased by political reasons.
Anyway, I will continue to publish my ideas through other channels, to the
global community. See you around
PS: my paper has an unconventional style which was *deliberate* to make it
more understandable.
--- REVIEW 1 -
SCORE: 1 (Weak Accept)
1. The paper is not in the required format.
2. The paper only described what is included in the proposed model, but
does not clearly explain how these parts work together as a complete
system, nor that why the system can be taken as an AGI.
3. Not sure how "inductive bias" is implemented in the system? Is it
"learned" by the system during the learning process, based on a control
mechanism of the system, or pre-defined when handling different problems
before the system starts learning?
4. "In principle, every state is potentially reachable from every other
state, if a logic rule exists between them. Now we use a deep FFNN to
represent the set of all logic rules." Theoretically speaking, yes, but are
inference rules between two states are handled continuously or discretely?
5. How interestingness is defined within the system? Why some propositions
have low interestingness and what is the purpose of having "forgetting
mechanism" in the addressed model?
6. Experiments or use cases implemented by the addressed model is preferred.
7. References from Wikipedia is not encouraged.
--- REVIEW 2 -
SCORE: -1 (Weak Reject)
There are a lot of promising statements in the abstract to the paper but
there is absolutely no justification of this statements in the main text.
Some of the ideas are looking plausible, but have no theoretical or
experimental verification. Some of the stetements are simply not correct,
e.g. statement about Turing completeness of RNNs (they really are but with
some additional refinement). Some of the statements are self-containing and
trivial. Also paper lacks of motivation for some of the presented models,
e.g. 15, which should be a picture (isn't it?). And two more major
drawbacks of the paper are its organization and appearance.
--- REVIEW 3 -
SCORE: -2 (Reject)
This paper is about inductive learning in the framework of AGI.
This paper does not contain any introduction and the reader is very quickly
a bit lost and does not know how to read and understand it.
The form of this paper should be entirely revised.
Then the paper is not easy to read and to follow, as it proposes an
unordered sequence of paragraphs with different topics and not necessarily
related.
At least, it is not indicated by the authors how we should read the paper
and what is the objective which is followed.
There is no much ore to say, this paper should be totally revised, on the
form and the content, to be at least readable and then evaluated in good
conditions.
--
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/T3cad55ae5144b323-M3e49379ebd86a9f57f10499b
Delivery options: https://agi.topicbox.com/groups/agi/subscription
