2017 54th ACM/EDAC/IEEE Design Automation Conference (DAC)
|When||Sun, Jun 18, 2017 — Thu, Jun 22, 2017|
Austin Convention Center
Austin, TX USA
|Topic||The Importance of Brain Theory in Creating Intelligent Systems|
The Importance of Brain Theory in Creating Intelligent Systems
Interest in deep learning systems is growing at a rapid pace. Yet, as experimentation with deep learning continues to grow, so does the awareness of its limitations. New methodologies and techniques are needed to create truly intelligent machines, and brain theory provides essential missing ingredients. Led by Jeff Hawkins, Numenta is on a mission to reverse engineer the neocortex, an approach that lays the groundwork for a new era of machine intelligence.
Numenta’s working computational theory of the neocortex is called Hierarchical Temporal Memory, or HTM. At the core of HTM are the time-based learning algorithms that store and recall spatial and temporal patterns. Unlike traditional machine learning techniques, HTM is distinguished in its ability to perform continuous online learning, handle multiple predictions, display robustness to sensor noise and fault tolerance, and perform well without the need for tuning.
In this talk, Numenta Senior Engineer Mr. Austin Marshall presents details and working principles of HTMs and raises important questions about the challenges and implications of this work. The following important questions will be discussed: How can we apply brain theory today? Where will it take us tomorrow? And what can we do to get there?
Specifically, Mr. Marshall will present recent research works by Jeff Hawkins and Numenta research engineers. An introduction to HTM and a theoretical framework for sequence learning in the cortex is presented. This research shows:
- How a biologically detailed model of pyramidal neurons with thousands of synapses and active dendrites can learn transitions of patterns.
- How a network of such neurons can form a powerful and robust sequence memory.
- How various properties of the model compare with several other neural network models.
Further, it is demonstrated how HTM can help us understand how the brain can solve sequence learning problems and how we can apply this understanding to real-world sequence learning problems with continuous data streams. This research shows:
- How various properties of HTM sequence memory are applied to sequence learning and prediction on streaming data.
- Comparisons between HTM sequence memory and other sequence learning algorithms.