Core Concepts
This paper proposes a new approach to Machine Learning that focuses on unsupervised continuous context-dependent learning of complex patterns, inspired by the structural and functional properties of the mammalian brain.
Abstract
The paper presents a new approach to Machine Learning inspired by the mammalian brain, particularly the cortex. The key ideas are:
Temporal Prediction based on Sequence Segmentation:
The central element is a Sequence Memory that identifies and labels segments (or sequences) in the input stream at each level of the hierarchy.
By segmenting the input stream, the system can handle very complex patterns with limited resources.
Short and Long Term Learning Modulations:
Learning is stochastic, with the probability of learning determined by the degree of knowledge of the current input sequence.
There are two states, Known and Unknown, that regulate the short-term probability of learning at the synaptic level.
Input Dimensionality Reduction, Pattern Disambiguation, and Feedback:
Each Sequence Memory handles a limited number of different values using k-winners take all inhibition.
Intracortical feedback helps to stabilize the inhibition processes and facilitate consistent symbol generation.
Spurious Identification Filtering, Learning Acceleration, and Learning Gating:
An auxiliary Sequence Memory, resembling the hippocampus, oversees the prevention of contamination of higher rungs of the hierarchy with spurious or incomplete identifications.
Corticothalamic loops are used to speculatively predict the input sequence identification of higher rungs, accelerating the identification process.
Lateral Contextualization:
Each cortical column has vertical inputs from a preceding column and lateral inputs from neighboring columns to generate a single identification modulated by the spatial context.
Supervised learning and Output Stabilization:
The system can incorporate mechanisms to direct the learning output of each cortical column to specific objectives and stabilize the output if it is in a long-term known state.
Brainstem and Context Stabilized Encoding:
The peripheral circuitry reduces the precision of the information supplied to the brainstem while maintaining the cortical disambiguation capabilities.
The proposed system is experimentally tested on the task of learning, identifying and predicting human speech, demonstrating its ability to learn with a reduced training set compared to conventional machine learning approaches.
Stats
This paper does not contain any specific metrics or figures to support the key logics. The focus is on describing the architectural design of the proposed Hierarchical Event Recaller system.
Quotes
There are no striking quotes in the content that support the key logics.