I wasn't impressed by current artificial neural networks because even a single neuron out of 14 billions in a human brain has to be much more intelligent than a few bytes of information and a threshold function. Some life forms have no more than that - a single neuron - still they manage to do so much. So perhaps a neuron has processing power comparable to that of a single modern PC. That processing power/memory is used for storing large internal statistics and plans on how to interact with the other neurons.
So let me present an attempt to implement a single neuron using all the resources of a single PC: https://github.com/srogatch/ProbQA
The project has both Applied AI and General AI goals.
In terms of Applied AI goals, it's an expert system. Specifically, it's a probabilistic question-answering system: the program asks, the users answer. The minimal goal of the program is to identify what the user needs (a target), even if the user is not aware of the existence of such thing/product/service. It is just a backend in C++. It's up to the others to implement front-ends for their needs. The backend can be applied to something like this http://en.akinator.com/ , or for selling products&services in some internet-shops (as a chat-bot helping users to determine what they need, even if they can't formulate the keywords or even their desires specifically).
The AGI goal is to achieve the processing power of a single neuron of a human brain on a single PC. Interactions with billions of other computers should achieve human-level intelligence (AGI). You can join the effort in designing how one neuron would answer the questions of another neuron, and how their collective intelligence rises from cooperation.
To avoid confusion with the current Artificial Neural Network terminology, we may have to call these units "artificial super-neurons".
Below are the learning curves of the program for matrix size 5 000 000: it's 1000 questions times 5 answer options for each question, times 1000 targets. In this experiment we train the program for binary search: the range of targets Tj is 0 to 999, and each question Qi is "How does your guess compare to Qi?". The answer options are 0 - "The guess is much lower than Qi", 1 - "The guess is a bit lower than Qi", 2 - "The guess exactly equals Qi", 3 - "The guess is a bit higher than Qi" and 4 - "The guess is much higher than Qi".
X-axis contains the number of questions asked&answered (up to 5 million). Y-axis contains for each 256 quizzes in a row the percentage of times the program correctly listed the guessed target among top 10 most probable targets. Note that testing is always on novel data: we first choose a random number, then let the program guess it by asking questions and getting answers from us, then either after the program has guessed correctly or asked more than 100 questions (meaning a failure), we teach the program, revealing it our selected random number.
From the data&diagram it seems that the program learns faster and reaches higher precision for priority functions which give more preference to lower-entropy options. So perhaps some exponential priority function can give superior results. But so far I don't know how to implement it without overflow. The priority function is in file ProbQA\ProbQA\PqaCore\CEEvalQsSubtaskConsider.cpp , near the end of it currently .
There is also a flaw currently in the key theory, which makes the program stubborn (I think it's close to "overfitting" term of Machine Learning). After the program mistakenly selects some target as the most probable, it start asking such questions which let it stick to its mistake, rather than questions which would let the program see that other targets are more probable. Although it is what happens in life, technically it is an error in the key algorithm/theory behind the program.
Current machine learning methods do not an learn algorithm. A classifier, a neural network or, more obviously, regression methods learn at most a function: given some input data, produce output data. That can be at most one step of an algorithm. Without learning algorithms, these methods are far from Artificial General Intelligence. On the path to AGI we should find a way for programs to learn algorithms with branching, loops and subroutines. Then we should find a way to teach programs to comprehend other programs. Then to comprehend and improve itself.
So in contrast to other machine learning methods, I'm currently developing and showing the first results of a computer program that learns algorithms. It has showed good results in learning the binary search algorithm (dichotomy). We can say that it learns an algorithm because what the program really does with its knowledge is interaction with the user in several steps: it asks the user for input, then depending on the input, takes some branch and asks the user something else.
In the previous articles I described a problem of the program bein stubborn: after convincing itself as if some wrong target is the most probable, the program was starting to ask questions which let it stick to its mistake, rather than learning the truth. I think I have solved this problem by changing the priority function from entropy-only based to distance&entropy based. Rather than just minimizing the entropy of posterior probabilities, it now also takes into account the Euclidean distance between prior and posterior probability vectors. This has allowed the program to learn 20 times faster. It now starts listing in top 10 the guessed target in almost 100% cases after about 125 000 of questions asked&answered (for a matrix of 1000 questions times 5 answers times 1000 targets). See the Accuracy graph below.
After the abovementioned initial training, it reaches positive result on average in 4.3 questions asked (by the program) and answered (by the user). See the graph below.
So that is quite competitive with the binary search algorithm programmed by a human. To narrow the search range from 1000 down to 10 targets, a human-programmed binary search algorithm would require 3.32 steps on average (it's base 4 logarithm of 100, because we have 5 answer options, of which one is just the strict equality).
However, human-programmed binary search algorithm doesn't tolerate mistakes and doesn't rate the targets by their probability of being the desired one. And of course it doesn't learn along the way. While probabilistic question-asking/answering system does this.
Moreover, I think there is some room for improvement in the priority function. Currently I use polynomial priority: pow(distance, 12) / pow(nExpectedTargets, 6), which empirically showed the best results in my limited number of priority function experimented with. Still I think that even better results can be produced with exponential priority function. Now let's see if it's better to devise and try some exponential functions now, or proceed with (less exciting) engineering tasks like saving&loading the knowledge base to&from a file, etc.
Recently the implementation of loading and saving of the knowledge base was finished. Training and prediction were finished more than a month ago, and were tested/fixed and tuned after that.
What's not yet finished is resizing the KB. It will not be possible to change the number of answer options after a KB is created. However, I am still to implement the changing of the number of questions and targets.
You can try integrating the engine into your systems. If not too busy, I may be able to answer your questions this integration.
To compile you need MSVS2017 v15.4.2 or higher. An external dependency is gtest: https://github.com/google/googletest (only if you want to run tests or dislike compilation errors in unrelated projects too much).
Earlier I published the results of experiments for top 10 targets (where a target is considered to have been guessed correctly if it's listed among 10 most probable targets). Here's the result for more challenging task - guessing the targets among top 1, i.e. it must be selected as the single most probable target. After many experiments and tuning, currently the learning curve of accuracy looks like the following for learning binary search algorithm:
So for KB of size 1000 questions by 5 answer options by 1000 targets, the accuracy reaches 100% after about 4.5 millions of questions answered, and then stays at 100% too.
More introduction to this program can be found in my articles: https://www.linkedin.com/in/sergerogatch/detail/recent-activity/posts/