**7. Summary and conclusion**

**Figure 13** summarizes our overall research vision and provides a broader context for our work. Current deep learning approach adopts the path on the left, employing high dimensional tensor processing and hardware accelerators with massive array of MAC units. Our approach takes the path on the right, leveraging spike timing processing and hardware fabric inspired by cortical columns. Our initial focus was on feed-forward mini-columns based on TNN principles. Our current focus broadens to

#### **Figure 13.**

*Summary of envisioned research in a broader context. Left path corresponds to the current deep learning approach that mainly employs numeric tensor processing deployed on hardware accelerators consisting of large MAC arrays. Right path illustrates our research approach that employs spike timing processing in the design and implementation of special-purpose processing units inspired by cortical columns, or Neuromorphic Intelligent Sensory Processing (NISP) units. Interesting cross-over ideas between both paths should be explored.*

all cortical columns with feedback in the form of reference frames. We are developing a framework and software tool suite for designing and implementing applicationspecific, highly energy-efficient sensory processing units, or Neuromorphic Intelligent Sensory Processing (NISP) units, with continuous online on-device learning capability. Our goal is to achieve potentially up to three orders of magnitude improvements on power and energy efficiency, relative to current deep learning accelerators. We also anticipate there could be interesting cross-over ideas between the current deep learning path and our new C3S path that should be explored.

This chapter presents a neuromorphic computer architecture and design approach that focuses on implementing neocortical computing fabric using digital off-the-shelf CMOS technology. This work builds on the foundational works of Jeff Hawkins' "*A Thousand Brains Theory*" and James E. Smith's *biologically plausible neural networks*. This research effort in NCAL at CMU aims to build Cortical Columns Computing Systems (C3S) that exhibit brain-like capabilities and brain-like efficiency. We hope our work serves as one step towards the holy grail of building a silicon neocortex. We hope to generate broad interest through this chapter that can lead to a vibrant research community pursuing this line of research. We believe there are tremendous opportunities for novel innovations that can have significant industry impact.
