One of my Research:
Breakthroughs in
|
inspired by Neuroscience.
|
inspired by Neuroscience.
Mimic birds revealed insights for explaining flight principles that led to the invention of airplanes.
As a Data Researcher, I mimic Computational Neuroscience hypotheses and apply scientific / engineering methods to present new solutions for cracking challenging AI/ML/Quant tasks.
As a Data Researcher, I mimic Computational Neuroscience hypotheses and apply scientific / engineering methods to present new solutions for cracking challenging AI/ML/Quant tasks.
Aiming to solve diverse Quantitative tasks focusing on delivering advanced algorithmic trading strategy for financial markets. Including data mining, asset allocation, risk management, robustness validation, etc.
Biological neural networks are very different from AI-based neuron networks. Still, the discovered paradigm shift revealed solutions to many classical AI-challenging tasks efficiently solved in the Brain. Moreover, revealed technological insights highlighted a way for AGI's practical implementation in a very appropriate environment of financial markets.
Biological neural networks are very different from AI-based neuron networks. Still, the discovered paradigm shift revealed solutions to many classical AI-challenging tasks efficiently solved in the Brain. Moreover, revealed technological insights highlighted a way for AGI's practical implementation in a very appropriate environment of financial markets.
A few of the Research models:
Novel AI Model
Novel AI Model to perform reinforcement, unsupervised continuous learning by exploring limited environment data and modeling the internal representation of the world. Able to generate a sequence of actions. Fast adaptation to changes in the environment.
Detailed study of the Brain learning methods by analyzing and modeling learning-related Neuroscience topics, including a few:
The discovered combination of unique bio-like topology and model training approaches, applied in deep learning, demonstrates a significant boost to the complex neural networks.
- Biological neurons learning approach;
- A critical function of apical dendrites;
- The role of Layer 1 in the neocortex;
- Basal Ganglia model;
The discovered combination of unique bio-like topology and model training approaches, applied in deep learning, demonstrates a significant boost to the complex neural networks.
The AI computational circuit comprises cortex areas, basal ganglia, Thalamus, hippocampus, input sensors, and actions space.
Cortex areas are made of cortical minicolumns (universal computational units of the Brain), which perform various tasks by sharing similar designs.
Researched a six-layers cortical column layout. Local and distant connections. Explained functionality, connectivity, properties, and types of neurons (pyramidal, VIP, SOM, PV, special functions) for each sub-layer. Computational algorithms and data structures for efficient processing of data. Model of cortical microcolumn development (growing).
Cortex areas are made of cortical minicolumns (universal computational units of the Brain), which perform various tasks by sharing similar designs.
Researched a six-layers cortical column layout. Local and distant connections. Explained functionality, connectivity, properties, and types of neurons (pyramidal, VIP, SOM, PV, special functions) for each sub-layer. Computational algorithms and data structures for efficient processing of data. Model of cortical microcolumn development (growing).
Model of Basal Ganglia, as a core part of AI system. The arrangement of the Basal Ganglia structural components for Reinforcement learning, actions execution, reward system, etc. World exploration, processing methods of complex patterns, data mining in Basal Ganglia. The rewarding system, behavioral and goal-directed acting.
Insights into Hippocampus data coding, storing world representation, and unique experience data saving.
An effective way to coordinate data flow and optimize computational power, implemented by concepts inspired by Thalamus.
Insights into Hippocampus data coding, storing world representation, and unique experience data saving.
An effective way to coordinate data flow and optimize computational power, implemented by concepts inspired by Thalamus.
Artificial bio-Neuron Model
AI neurons are significantly different from Biological neurons. Developed models of bio-Neuron open new possibilities to tackle practical problems in entirely new ways.
The AI neuron computes the output value as a weighted summation of inputs transformed by an activation function. Backpropagating prediction errors update weight parameters.
Today's AI has accomplished many remarkable tasks. But AI neurons and deep neural networks have problems like gradient vanishing, fixed topology, long and expensive training, large training data sets, etc.
The insights received from a detailed analysis of biological approaches make it possible to rethink the practical implementation of Deep-learning by redesigning the calculation flow. Aiming to amplify the advantages of both appraoches.
The AI neuron computes the output value as a weighted summation of inputs transformed by an activation function. Backpropagating prediction errors update weight parameters.
Today's AI has accomplished many remarkable tasks. But AI neurons and deep neural networks have problems like gradient vanishing, fixed topology, long and expensive training, large training data sets, etc.
The insights received from a detailed analysis of biological approaches make it possible to rethink the practical implementation of Deep-learning by redesigning the calculation flow. Aiming to amplify the advantages of both appraoches.
Bio-Neuron parts:
Each part is made of building segments.
Segment described as:
Particular Nodes units are modeled by:
By changing the parameters of nodes, bio-neurons can perform very different functions, generate various output signals, and describe unique neuron properties. (like excretory neurons such as Pyramidal or Inhibitory such as PV, SOM, VIP, ...)
- Dendrites - receive inputs signal;
- Soma - generate output signal;
- Axon - delivers an output signal to other neurons;
- Apical Dendrites - perform specific functions, including network learning;
Each part is made of building segments.
Segment described as:
- Connectivity table with the values of linked segments ID, conductance, latency, and other parameter values.
- Segment properties, like a capacitor, voltage, and other specification values.
- Set of Particular Nodes in the current segment.
Particular Nodes units are modeled by:
- parameters values;
- finite state space stochastic model;
- state transitions events, which could depend on specific parameter values (e.g., voltage or concentration of neurotransmitters.)
- Ion channels (K+, Ca2+, Na+, Cl-, etc.)
- Synapses (like AMPA, NMDA, GABA, ...)
By changing the parameters of nodes, bio-neurons can perform very different functions, generate various output signals, and describe unique neuron properties. (like excretory neurons such as Pyramidal or Inhibitory such as PV, SOM, VIP, ...)
Discovered a special rule for the Bio-Neuron Learning algorithm, different from Hebbian Learning or other types of plasticity. Moreover, the rule that makes practical engineering meaning.
The single algorithm is used for updating parameters of:
Bio-Neuron model design makes it possible to change functions by updating nodes and segment parameters by the Learning algorithm. Moreover, it could grow or prune its parts with the same unique model to establish new connections with other neurons or move in space. Flexible topology made possible.
The single algorithm is used for updating parameters of:
- nodes;
- parts;
- segments;
- growing on a neuron;
- signaling;
- synapses strength modification;
- ...
Bio-Neuron model design makes it possible to change functions by updating nodes and segment parameters by the Learning algorithm. Moreover, it could grow or prune its parts with the same unique model to establish new connections with other neurons or move in space. Flexible topology made possible.
A detailed understanding of Bio-neurons properties makes it possible to creatively adapt it in designing computational units in Deep-learning.
Still, it's a topic of research to which extent it's the most efficient to repeat biological neurons, but I believe the most efficient design is somewhere in the middle of Bio and ANN neurons. Moreover, the Bio-like neurons represented a specific architecture of DNN, which exploit the discovered hints from the Brain-like design. Interesting practical observations, that the same computational principles observed on a different scale of Brain works.
Still, it's a topic of research to which extent it's the most efficient to repeat biological neurons, but I believe the most efficient design is somewhere in the middle of Bio and ANN neurons. Moreover, the Bio-like neurons represented a specific architecture of DNN, which exploit the discovered hints from the Brain-like design. Interesting practical observations, that the same computational principles observed on a different scale of Brain works.
Evolutionary Solutions
Self-improvement of AI-model blueprints by evolving logic architecture powered by genetic algorithms.
A predefined genetic-like code data structure describes component parameters and structural architecture. Practical usage in storing networks and distributed computation parameter synchronizations.
Developed algorithms for mutation, crossover, and selection operations on multiple levels of the AI system.
A solution for direct estimation of various components' optimal parameter values.
Developed algorithms for mutation, crossover, and selection operations on multiple levels of the AI system.
A solution for direct estimation of various components' optimal parameter values.
Lifestages models for Artificial species as a consequence of genetic code for component description. Genetic code structure for enhanced AI systems learning in a massive neural network.
Deployed method for representation of species by a picture with color spots (e.g., matrix code), which is updated by the genetic operations. Used as a core concept for evolving cortex area, responsible for modeling other species' behavior (inspired by face recognition) and boosting better AI systems.
Offered the next level of environmental complexity by adding reproduction elements, which require a partner selection. Enhance species traits for specialization architecture. Enhancing specialization by adding features for cooperative working.
Presented particular species, focusing on understanding genetic code and performing reproduction guiding to discover better architecture.
Deployed method for representation of species by a picture with color spots (e.g., matrix code), which is updated by the genetic operations. Used as a core concept for evolving cortex area, responsible for modeling other species' behavior (inspired by face recognition) and boosting better AI systems.
Offered the next level of environmental complexity by adding reproduction elements, which require a partner selection. Enhance species traits for specialization architecture. Enhancing specialization by adding features for cooperative working.
Presented particular species, focusing on understanding genetic code and performing reproduction guiding to discover better architecture.