Models of WorldBrain
WorldBrain can be quickly trained in an unsupervised manner to learn compressed spatial and temporal representations of the environment. By using features extracted from the world model as inputs for the agent, a very compact and simple strategy can be trained to solve the required tasks. The strategy can even be trained entirely within the illusions and dreams generated by its world model and then transferred back to the actual environment.
Agent Model
The purpose of agent model is to find a balance between performance and efficiency. In this model, our agent has a visual sensory component that compresses what it sees into a small representative code. It also has a memory component that makes predictions about future codes based on historical information. Finally, our agent has a decision-making component that decides what actions to take based only on the representations created by its vision and memory components.
Controller (C) Model
The WorldBrain Controller (C) model is primarily used to implement the control and decision-making functions of the WorldBrain system. The controller is a core component of the WorldBrain system, responsible for receiving input information, processing and analyzing data, and generating corresponding outputs and decisions. The design goal of the Controller (C) model is to simulate the control system in the human brain to achieve intelligent and autonomous behavior.
MDN-RNN Model
The MDN-RNN (Mixture Density Network - Recurrent Neural Network) model combines a mixture density network. with a recurrent neural network. It is employed for modeling multi-modal outputs in sequential data. This model finds broad usage in generating and predicting sequential data, such as music generation and text generation.
VAE (V) Model
VAE（Variational Autoencoder）is a generative model that involves the process of encoding (Encoder), decoding (Decoder), and inferring the latent space. It is utilized to learn and generate data by modeling the underlying distribution of the data. VAE has widespread applications in tasks like image generation and data synthesis.
Storage Mechanism in WorldBrain Models
Knowledge in the brain is distributed storage. All knowledge is not confined to a single location, such as a single cell or cortical column, nor is it stored like a hologram where everything is stored at any one place. Knowledge about an object is distributed across thousands of cortical columns, and neurons never rely on individual synapses. This concept is a fundamental aspect of WorldBrain development. In the WorldBrain simulation network, a decentralized storage approach is employed, and even if 30% of neurons are lost, the impact on network functionality is usually minimal.
Voting Mechanism in WorldBrain Models
Perception in the brain is a consensus achieved through voting in cortical columns. The majority of connections within the cortical columns move between layers, mainly residing within the boundaries of the cortical column. Cells in certain layers send axons to very distant regions within new cortex, potentially crossing from one side of the brain to the other, for example, between regions representing the left and right hands, or they may send axons from the primary visual area V1 to the primary auditory area A1. Cells with these long-distance connections participate in the voting process.
Prediction Mechanism in WorldBrain Models
WorldBrain simulates the brain using a "memory-prediction model," where the new cortex learns a World Model and makes predictions based on that model. The brain creates a predictive model, signifying that the brain continuously predicts what the input information will be. Prediction is not something the brain does all the time but is an inherent property of the brain. The brain never stops predicting, and prediction plays a crucial role in the brain's learning. When the brain's predictions are validated, it indicates the accuracy of the World Model within the brain.
Neural Network Simulation
A useful action from a neuron takes at least 5 milliseconds. The speed of silicon transistors is a million times faster than neurons. Therefore, a brain neocortex made of silicon could potentially be a million times faster than human thinking and learning. WorldBrain can utilize neural network models to simulate neurons and the neural network structures found in the old brain system. By constructing deep neural networks, WorldBrain can replicate the information processing and transmission of the old brain system.
These neural network models may include perceptual networks, memory networks, decision networks, and more, in order to mimic the different functions and features of the old brain system.
World Model is the
future of AI
World Model is the advanced model in the field of artificial intelligence, commonly used in reinforcement learning. It constructs and predicts the dynamic characteristics of the environment, enabling AI to simulate and predict future states to assist in decision-making. World Model can be seen as the "mental model" of an AI system, reflecting the system's perception and expectations of itself and the external world.
The future development of artificial intelligence will involve more technologies and methods, including more powerful neural networks, advanced self-learning capabilities, and more sophisticated perception and reasoning techniques. World Model has been achieving success in exploring multi-task learning and environmental modeling. It must be the ultimate path to achieve AGI.