Biomorphic AI

Biomorphic AI

It’s known that a man has created many algorithms based on nature laws: genetic algorithms, algorithms of neural networks, swarm intelligence, etc. The most effective methods of AI are effective simplification and modeling of processes in nature. But, unfortunately, experts are often well versed in only a few models, when they read information from other articles they feel dizzy and they still remember only what is related to their topics. In order for everyone to be able to discover such a complex topic as biomorphic AI, we offer a hierarchy format: topic → article → application → pros and cons. To demonstrate how this works, we will show a description of the positive applications of linking algorithms with their biological roots.
Biomorphic AI
Application pros cons
Consideration of several options for computational simplification of the original algorithm in compliance with their biological basis Potential efficiency gains Narrowing of scope is possible
Consideration of several options for how to make the algorithm more suitable for the biological basis ability to model some biological processes; Improving the adaptability of the algorithm to “natural” data Potential deterioration in efficiency

Smart technologies and recreation of biological processes

Dropout

Dropout
Application pros cons
Dropout 20% - 30% of neurons Efficiency rarely decreases and often rises With strong overfitting, it may not give an optimal result
Dropout 30% - 50% of neurons Strong overfitting can improve efficiency Probability of a slight decrease in efficiency
Many stable natural systems are based on unstable by themself elements. One of these systems is our brain and its work is based on the transmission of impulses between neurons. In order to adapt such systems, a dropout algorithm was invented - the accidental deletion of neurons in an artificial neural network or elements in any other similar system. It helps to make the system more stable and, in the case of artificial neural networks, reduces overfitting with available data. The authors of this article compare the loss of neurons with the transfer of genes during sexual reproduction, as in reproduction only a certain part of genes is realized in the offspring. Let’s suppose that an individual had good traits, but these traits were determined by a large set of genes, or, in other words, a whole large coadaptation of genes. Then this complete set of genes, most likely, will not be passed on to the next genera. For generations, the offspring would undoubtedly contain only genes that play a positive role for humans, or small useful groups of genes.
Additional literature:

Small batch training

Small_Batch
Application pros cons
A selection of 2-4 examples Sometimes it can be more generalized; Sustainable Performance Compared to Other Use Cases Greater increase in time per epoch; The risk of getting an unsatisfactory result
A selection of 4-16 examples Usually has an optimum in learning time to convergence and an optimum in efficiency improvement A slight increase in the time spent per era
A selection of 16+ examples Less time per epoch; Sometimes, with not very sensitive data, it can give an optimal result less stable result
Artificial neural networks, unlike humans, have the ability to study all various information simultaneously, without intermediate performance tests. A person cannot do this, as he needs to apply the information he receives almost immediately. But this turns out to be even not bad: if the average indicators in a piece of information (sample) have small deviations from the global one, then they will not allow leaving the area of the global minimum error (corresponding to the complete information), but at the same time they will help to leave the area of local minima(which appear due to the presence of special noise in the input data). This approach is called Small Batch Training.
For neural networks, it is optimal to choose from m = 2 to m = 32.
Small Batch Training can be seen in all of our training:
organizing lessons in schools; lectures and practices at universities; online courses;
This one is very useful when there are random collective correlations in the data.

Fast R-CNN

Fast_R-CNN
general cons: not the fastest method, complex architecture (more difficult to train together with other neural networks)
Application pluses cons
Localization of small objects works well enough
Localization of Medium Objects works well enough
Determining the boundaries of an object works well enough
If we compare the neurophysiological structure of human hearing and vision with neural networks, then we can say that hearing is a sequential model of neural layers, which leads from sensors to the perception of such high concepts as, for example, musical harmony, on the other hand, information gets into the eyes and undergoes preliminary processing for the simplest signs begins to be further processed by two other neural networks - one which is responsible for determining what kind of objects in the picture, and the second determines where exactly these objects are located. And now, moving away from the comparison and thinking about what is written, we see that this is fully consistent with Fast R-CNN.

Faster R-CNN

Faster_R-CNN.bmp
general cons: slightly complex architecture (slightly more difficult to train with other neural networks)
Application pluses cons
Localization of small objects does not work very well
Localization of Medium Objects works well enough
Determining the boundaries of an object does not work very well
When a person is focused or wants to quickly perceive the situation, he tries to peer into each changed piece of the image and quickly assess the position of new objects piece by piece and, having put the picture together, understand what is happening. This approach is closer to Faster R-CNN.

Recursive Cortical Network

Recursive_Cortical_Network
Application pluses cons
Defining an Object Hierarchy works well enough works too long because of what it is ineffective to apply to real images
Object recognition from the object system works well enough works only for a system of simple objects
The algorithm, unlike its predecessors, does not decrease the accuracy when deforming the text and almost does not decrease the accuracy when using various effects and styles to complicate the text and overlap it with other objects.
Thanks to him, he allowed to correctly guess Captha in 66.6% of cases, and after determining the style and additional training - in 90% of cases. Moreover, this network uses only 5000 examples of solved captcha examples and a small number of layers. Simulates the work of the primary visual cortex. A description of the work and experiments with this neural network is given in this article. Uses lateral communication between neurons to reveal more flexible patterns in information.
Additional literature:

Basic algorithms for a potential strong AI

AlphaZero

AlphaGo_Zero
Application pluses cons
Self-study of techniques in board games brings out new playing techniques that people can learn from The Difficulty of Perception Techniques for Non-Professionals
Learning Games That Simulate Utility Models can provide new methods of competition and promotion methods are difficult to fully understand
Games wins Experience for world champions - competition with a very strong opponent, entertainment in the early stages After constant victories AI entertainment falls
In the beginning, there were algorithms that, according to some criteria, assessed the effectiveness of certain moves and thus determined the optimal one. To improve this estimate, a Monte Carlo tree algorithm was made - the algorithm, choosing randomly (but with preference for more optimal options), simulated several games by itself simultaneously determining which move would lead to a more optimal future and gradually building a tree from simple estimates of game positions (thus simulating human-like planning), this approach peaked at the level of the average go-lover.
Monte_Carlo_Tree_Searh
Replacing the assessment of the position, which was made on the basis of the subjective understanding of Go players, with the assessment that is made using a convolutional neural network (or rather, a residual neural network) in DeepMind, they received AlphaGo, which later defeated the best player in Go.
Further improvement AlphaGo Zero, learned to learn not only on external data, but also playing with itself. This not only got rid of her dependence on external data, but also significantly improved the program.
AlphaZero is a program that, learning from scratch, conquered all people and other algorithms in chess, shogi and go. AlphaZero can learn to play any game and become the best at it in a matter of hours. It adapts itself to the game, which makes it quite flexible and easily customizable for different types of games.

BELBIC (Brain Emotional Learning Based Intelligent Controller)

BELBIC
Application pluses cons
Simulating Living Creatures Good for simulating a small group of simple animals describes the behavior of smart animals with difficulty; computationally difficult to simulate a large number of individuals
Emotion Modeling Simulates the emotional reactions of animals and humans well does not include planning for animals, let alone consciousness
Dynamic adaptation Computationally small model The need to combine with more stable algorithms
Such mechanisms of our psyche as attention and emotions help us highlight the most important information in order to quickly analyze something or not overstrain us. Emotional brain training models are called BEL (Brain Emotional Learning) models. The most advanced of these models is the BELBIC model, Such models themselves are interesting in that they simulate some part of the human limbic system and therefore can potentially have some similar internal rhythms, etc., but besides this, it is also a computationally efficient model for regulating the parameters of your control of drones and other equipment.
Additional literature:

MuZero

MuZero
general cons: unstable training, in order to bring the level to the highest level, it is necessary to have a set of neural networks-players of an increasing level of complexity in order for the neural network itself to be able to gradually win one by one to reach the peak of a possible level
Application pluses cons
Self-study of techniques in any games brings out new playing techniques that people can learn from The Difficulty of Perception of Techniques for Non-Professionals
Use to obtain new drugs in medicine high speed of work The need for security and risk control
Use of computer or board games Experience for world champions - competition with a very strong opponent, entertainment in the early stages After constant victories AI entertainment falls
Simulation of Human Behavior Teaches planning and learns to properly simplify your perception The peculiarities of human emotions and behavior are not taken into account
Simulating Other Utility Models can provide new methods of competition and promotion methods are difficult to fully understand
In order to understand what MuZero is, read about AlphaZero. MuZero, unlike AlphaZero, is not tied to what data is submitted for input, MuZero must learn to compress information himself, this information is, relatively speaking, a subjective vision of reality by an algorithm. This view of reality be in the hidden state - we can say that this is the digital equivalent of the hippocampus in humans.
Additional literature:

Multi-task learning

Hierarchical Regularization Cascade

Application pluses cons
Defining a Property Hierarchy prevents retraining, highlights the system in information a little more time, the need for a preliminary initial grouping of tasks, the subjectivity of the allocated hierarchy of relations
Mutual improvement of the qualities of existing neural networks Improved stability, potential efficiency gains the need to find a neural network training on similar data
In our life there are many areas and sub-areas of various tasks, some of these divisions are conditional, and some are not. This algorithm tries to gradually select useful information on the basis of information, and it learns to understand where and in what area or task this or that information is useful, thereby reducing retraining and increasing the stability of the algorithm.

Convex Feature Learning

Application pluses cons
Defining a Property Hierarchy prevents retraining, completely independently selects the system in the information more time, the subjectivity of the distinguished hierarchy of relations
Separating Features in a Hierarchy the presence of a clear graph indicating the hierarchy; The ability to effectively retrain the algorithm for a new distraction without rebuilding the hierarchy Complexity of Effective Extension of the Algorithm for a Large Number of New Problems
Sometimes there is no clear hierarchy in task topics, but there are interdependencies. Trying to search for them and highlighting the graph, you can get your own subjective and effective hierarchy of task topics.
Additional literature:

Exploiting Unrelated Tasks

Unrelated_Tasks
Application pluses cons
Mutual improvement of the qualities of existing neural networks Improved stability, potential efficiency gains Due to penalties for similarly solving problems from different groups, strong interrelationships between groups of problems can disrupt the algorithm
Sometimes in life we have to take on incoherent tasks, but by learning to plan and competently separate and organize our mental processes, we can potentially learn to solve our old tasks even more stably and efficiently.