When David Sherwood started college as a physics major, he had one question that had nothing to do with physics: How does biological memory work?
It was obvious, even then, that a photograph remembers what it sees because of a chemical reaction. But what was happening inside the brain? He read every book on neurophysiology he could find in the library, landed on one theory he thought was close but not quite right, and started building his own.
That was decades ago. He’s still building it.
The career that followed kept circling back to the same set of problems. At AEL in Lansdale, Pennsylvania, Sherwood worked as a systems engineer on an airborne surveillance system, learning radio technology and phase interferometry. At ITT Avionics, he designed the first all-microprocessor electronic countermeasures system for the SR-71. At Litton AMECOM in College Park, Maryland, he ran integration and testing on a submarine-based electronic warfare system, an experience he credits with teaching him more about how not to design computer systems than how to design them.
Then came RCA. In 1980, the Army’s Electronic Warfare Command needed a computer small enough to fit on a helicopter that could process two million radar pulses per second, detecting and sorting signals fast enough to protect the aircraft from radar-guided missiles. Sherwood was given the job of designing it.
He solved the problem using an unconventional approach. When RCA’s own experts reviewed the design, their verdict was definitive.
“A, it can’t be built. B, if we could build it, it wouldn’t work.”
There were personnel changes. A new manager built it. It worked exactly as intended.
A year later, Sherwood proposed using neural networks for electronic warfare applications. The Army Ft. Monmouth engineers he first met with loved the idea. The most senior person in the chain did not.
“No, neural networks are dead,” Sherwood recalls being told. That period is now known as the AI winter.
Through all of it, the memory question never went away. As he gained engineering skills and better mathematical tools, each new piece of knowledge fed back into the theory he’d started sketching in college.
The mainstream theory of memory, first articulated by Donald Hebb in 1949, holds that memories form by adjusting the strength of existing connections between neurons. That principle underpins virtually all modern AI. Every large language model, every generative AI system, learns through a process called backpropagation: tuning the weights between artificial neurons, nudging them up or down until the system gets better at matching inputs to outputs.
Sherwood’s theory says something different. Instead of changing the strength of existing connections, his model creates new ones. When the system learns something, neurons that weren’t previously connected suddenly are, and their weights are binary: zero or one.
“Conceptually, that’s not a big difference,” Sherwood says. “But in terms of computers, it’s an enormous difference.”
Because all the connection weights in his system are zeros and ones, the system doesn’t need to multiply anything. Traditional AI’s enormous energy demands come largely from multiplying fractional weights across billions of connections. Remove the multipliers, replace them with simple logic gates, and the power requirement drops dramatically.
His system works more like a database running on neural networks. Every row of data lives in its own separate neural network. You can read what’s stored, inspect it, fix errors, and predict the system’s behavior. There’s no training phase, no black box.
“When you’re done, and you’ve trained millions, perhaps billions of neurons and synaptic connections with all of these weights, you cannot look at the billions of connection weights and say, ‘What does it know? What did it learn from the training?’” he says of conventional AI. His system, by contrast, has the transparency and controllability of a traditional database with the generalization abilities of a neural network: give it an input, and it finds the best matching output without requiring an exact-match search.
Sherwood started working on AI before ChatGPT and large language models arrived. When they did, the contrast with his own approach sharpened. Today, Cognitive Science & Solutions is based in Eldersburg, Maryland, half an hour west of Baltimore, in a community that exists almost entirely to serve the intelligence agencies nearby. Sherwood spent more than two decades as a lead systems engineer in electronic warfare, signals intelligence, and information operations, working on systems for the NSA, CIA, Army, Navy and Air Force.
A few years ago, his team was awarded a patent on a new computing chip they call Hercules ER.
“It’s Hercules because it’s so freaking powerful, but it requires almost no energy,” he says.
The chip has no multipliers. What it has is hundreds of thousands of AND gates, each one roughly four transistors, processing enormous amounts of data in parallel. A second patent is in progress, covering how to systematically represent database information within a neural network. And the company’s Diamond AI standard platform consolidates years of one-off demonstrations into a single standardized architecture that can be customized for specific problems or licensed directly to customers who want to build their own solutions.
Sherwood is the first to acknowledge what’s missing. Internal demonstrations have proven the technology works, but the outside world wants independent verification.
“That’s probably been the biggest weakness we have,” he says. “We’ve proven to ourselves that it works, but that’s not good enough for the outside world.”
The company is now pursuing collaborations with universities and other organizations to provide that independent testing.
He’s also candid about his own blind spots. A career spent entirely inside the Defense Department left him with deep expertise in DoD applications and limited experience in the commercial world. His solution has been to surround himself with people whose strengths complement his own.
“I’m very good at understanding how to apply this technology to a lot of DoD type applications,” he says. “That’s a piece of cake for me. I’m a little bit blindsided outside of that.”
When it comes to advice for other innovators, Sherwood draws a sharp line between two types of ventures. If your innovation is an improvement on existing technology, something investors and customers can easily understand, go for it. If you’re doing something genuinely new, prepare for resistance.
“I have been told over and over again that my battle’s gonna be tough simply because there is so much groupthink,” he says. “When people hear I don’t use backpropagation, their minds immediately turn off.”
Outside of work, Sherwood loves sports. He played touch football every Sunday for years until too many of the other players moved away. He’s a tennis player, a former scuba diving instructor, and someone who would happily get back in the water if the chance came around again.
The next milestone for Cognitive Science & Solutions is to find one solution so strong it rivals or beats anything else on the market, then prove it with independent verification. Not a dozen demos across different domains, but one application built with enough depth to make the case on its own.
“It’s hard coming in with something that’s really different,” he says. “Which is why it’s taken me decades to make much progress.”
He’s been hearing that his ideas can’t be built, won’t work, and are dead since 1980. He’s still building them.
David Sherwood is the Founder of Cognitive Science & Solutions, based in Marriottsville, Maryland. A career systems engineer, he spent more than two decades working in electronic warfare, signals intelligence, and information operations for the defense and intelligence communities. His early work includes designing the first all-microprocessor electronic countermeasures system for the SR-71. Cognitive Science & Solutions has developed a patented approach to artificial intelligence that uses binary-weight neural networks, eliminating the need for backpropagation and dramatically reducing energy requirements. The company holds a patent on the Hercules computing chip and is developing its Diamond AI standard platform for commercial and government applications.