As researchers strive to create artificial intelligence that rivals the capabilities of the human brain, a key challenge has been translating the brain’s intricate synaptic connections and neural dynamics into electronic hardware. However, a novel platform called Cog Ex Machina is making important progress in bridging the gap between biology and silicon.
Developed through a collaboration between researchers at Hewlett-Packard Laboratories and Boston University, Cog Ex Machina is a hardware and software architecture designed to enable large-scale brain modeling and neuromorphic computing. At the heart of the platform are memristive devices – electronic components that can change resistance based on voltage history, similar to how synapses in the brain adapt.
By integrating dense, low-power memristive memories directly with computational circuitry, Cog Ex Machina aims to mimic the brain’s efficient information processing. The platform’s digital foundation provides flexibility, allowing researchers to implement a wide variety of neural network algorithms and learning rules.
“Cog effectively abstracts away the details of the underlying hardware, so researchers can continue building models on the platform as the technology advances,” explains Greg Snider, a senior researcher at HP Labs.
Indeed, the software framework called Cog, which runs on top of the Cog Ex Machina hardware, provides researchers with a modular, graph-based approach to building large-scale brain models. Computational nodes in the graph implement adaptive transformations, ranging from linear operations akin to dendritic processing to nonlinear dynamics representing neural population responses. These nodes communicate through tensor fields, analogous to the information conveyed by axon bundles.
To test their brain models, researchers can plug in “animats” – either virtual software agents or physical robots – that allow the models to interact with simulated or real-world environments. This enables studying how the brain models learn and adapt, rather than just assessing their static behavior.
The Cog Ex Machina team has already demonstrated several proof-of-concept applications, from contrast normalization inspired by retinal processing to the self-organization of orientation and ocular dominance maps in the visual cortex. Importantly, the platform’s flexible software architecture and rapid development cycle allow researchers to quickly explore new neural algorithms and hardware configurations.
“Cog has many features that offer this flexibility,” notes Massimiliano Versace, director of Boston University’s Neuromorphics Lab. “Its all-digital hardware foundation reduces technological and fabrication risk, while the tensor framework mechanisms pull in a wealth of mathematical and engineering knowledge to enable the exploitation of powerful computational techniques.”
As the underlying memristive hardware and neural modeling techniques continue to advance, the Cog Ex Machina platform represents an important milestone on the path toward creating artificial brains that can rival the flexibility, efficiency, and capability of their biological counterparts. By bringing together the latest developments in neuroscience, computer science, and electrical engineering, this collaborative effort is pushing the boundaries of what’s possible in brain-inspired computing.