Trainable personal robots that behave and learn naturally, inspired by neuroscience, are now one step closer to reality with the development of an internal representation system based on Hierarchical Temporal Memory (HTM). HTM, a computational theory capturing core algorithms used by the laminar neocortex, enables robots to build and leverage a learned model of reality, similar to how animal brains function.
The key to replicating a basic brain lies in defining the scope of what brains do. At its core, a brain learns models and makes predictions about its senses and environment to facilitate reasoned actions. This ability to represent reality internally is crucial for context-aware decisions, which is something all robots must possess in order to respond intelligently based on the situation.
The cortex, present in all living things, is responsible for representing the environment internally. It contains hierarchical networks that learn models of senses across time and space, allowing for an understanding of reality. It can be seen as a common subspace for all sensors, integrating different sensory inputs into a unified framework of knowledge about time and space. This convergence enables inference and prediction.
HTM theory encapsulates the algorithms employed by the cortex, including spatial pooling, temporal memory, and sensor fusion. By implementing an HTM framework, robots can store representations of the environment and generate meaningful responses, similar to animal brains. This approach offers several advantages over conventional programming, such as producing inherent invariances automatically from real-time data, allowing high-level inferences based on sequenced low-level observations, and handling novel inputs more effectively.
The HTM cortical learning algorithms consist of three core functions. Spatial pooling performs pattern recognition on input data to identify common features, regardless of location. This mechanism is analogous to visual edge detectors. Temporal memory models time, discovering causes and effects by recognizing spatial patterns in sequence. It can be considered analogous to grid cells activating at specific moments. Sensor fusion integrates different sensory inputs spatially and temporally, simulating the experience of hearing a bark and seeing a dog simultaneously from different senses.
By building hierarchical networks using these principles, HTM can capture increasingly sophisticated features and behaviors over broader timescales. This replication of cortical structure allows HTM-based robots to display more adaptive responses, similar to animals. Biologically inspired robotics through the exploration of this potent concept is the aim of this project.
With its potential to revolutionize the field of robotics, HTM opens up new avenues for creating trainable personal robots that can behave and learn naturally. By emulating the core algorithms of the cortex, these robots can form internal representations of their environment, enabling them to make intelligent decisions based on context. As the development of HTM-based robots progresses, we can expect to see increasingly adaptive and human-like interactions between man and machine.
