Biographies · Neuromorphic engineering and sensory computation
In April 2024 Carver Mead stood up in La Jolla to accept an award for helping to found neuromorphic engineering. The prize bore the name of his former student Misha Mahowald. When it was his turn to speak, Mead told the audience who had first wanted to build a retina out of silicon. It was Mahowald, he said, who had drawn him into neurobiology. They had made a good pair. She wanted to understand how an eye works. He had spent years teaching people how to turn an idea into a chip they could hold and test. Mead’s recollection at the award ceremony.
Importance of his work for artificial intelligence
Draw a thick pencil line on a sheet of white paper and carry it from a bright window into the shade. The paper now sends less light to your eyes. So does the pencil mark. Yet the edge of the line is still easy to find. How could a machine pick it out in both places? One clue is the difference between neighbouring spots. A spot in the middle of the paper has neighbours much like itself. A spot beside the pencil line has a dark neighbour on one side and pale paper on the other.
On Mead and Mahowald’s chip, each light-sensitive element produced a signal, and its neighbours supplied a local reference. The wiring made the comparison as the light fell on the silicon. Imagine using those signals to read a handwritten P. They could help you trace its upright stroke and its loop. An R has both as well. To tell them apart, a reader would also have to find the slanting leg. The chip supplied evidence that a recognition system could use. Its designers had built this first step of vision into the sensor itself. Mead and Mahowald’s retina experiment.
A house beside the generators
Mead grew up at Big Creek in California’s Sierra Nevada, where his father worked at a hydroelectric plant. In his Kyoto Prize lecture he showed the audience where his family had lived. In the same photograph he could point out the power station and the little school up the hill. One teacher taught eight grades in a single room. The station had instruments that let its operators follow the behaviour of enormous generators. Long before Mead entered a laboratory, he had seen what it meant to make a physical process give a reading that someone could act on. Mead’s illustrated lecture · His Caltech oral history.
He came to Caltech in 1952. He earned his bachelor’s degree in 1956, his master’s in 1957 and his doctorate in 1960, all in electrical engineering. By the time he finished the doctorate he was already teaching. The transistor gave him a device small enough to study at the level of its materials, yet versatile enough to become part of an entire machine. The difficulty for a student with a promising design was getting it made. Caltech’s biographical record · Mead at Caltech.
Making room on a wafer
In his microelectronics course, several students’ designs could share a single silicon wafer. The factory could make them in one run, and each student could test a chip of their own. With Lynn Conway, Mead helped turn chip design into a subject that a new generation of engineers could learn. Their book Introduction to VLSI Systems explained how to lay out complex circuits using design rules. Following those rules allowed a designer to work without mastering every chemical operation in the factory. VLSI stands for very-large-scale integration, the art of building many circuit elements together on a chip. The Kyoto Prize account of Mead and Conway’s work.
What counts as seeing?
The retina at the back of an eye does some work before a signal ever reaches the brain. That made it an inviting subject for an experiment. In 1988 Mead and Mahowald reported measurements from an array of 48 by 48 picture elements. Each element responded to light on a logarithmic scale, which allowed it to represent a wide range of brightness. A network of electrical connections supplied a local average. The difference between the individual reading and that average brought out changes across the scene. The authors compared signals from their chip with responses of bipolar cells, one kind of cell in a vertebrate retina. That comparison gave them something more demanding to test than a visual resemblance between two pictures. Mead and Mahowald’s measurements.
Mahowald went on to ask what a pair of these devices might do together. Her 1992 doctoral thesis described a system that took signals from two silicon retinas and matched features between the two views to estimate depth. She also developed a way for chips to exchange signals inspired by the impulses that travel along nerves. The original experiment, the doctoral work and Mead’s later account tell different parts of the story. In the thesis it is Mahowald herself who sets out the questions and explains how she tested the answers. Mahowald’s doctoral thesis.
A piece of silicon that sorts sounds
With Richard Lyon, Mead turned to the cochlea, the spiral structure in the inner ear. Imagine striking a piano key. The vibration that enters the ear sets a wave travelling along the cochlea, and where that wave reaches its peak depends on frequency. In their 1988 circuit, a succession of electronic filters played the part of successive stretches of the cochlea. Measuring the signal at different stages showed how the circuit responded to different tones. Sorting sound by frequency is a useful first step for a device that listens. Understanding a spoken word is a further problem. Lyon and Mead’s test chip.
Mead brought these experiments into the classroom and into his 1989 book Analog VLSI and Neural Systems. In 1990 he set out the reasoning in a paper called Neuromorphic Electronic Systems. In an analogue circuit, a current or voltage can vary continuously with the quantity it represents. Connecting devices in a particular way can make the currents themselves carry out a comparison. It also means that small differences between real components matter. How the circuit adapts to those differences was part of Mead’s argument from the outset. Mead’s 1990 paper · His account of the experiments and the book.
Which way did the sound come from?
Close your eyes and ask someone on your left to clap. The sound reaches your left ear a little earlier than your right. That small lead is a clue to direction. In 1989 John Lazzaro and Mead described a chip modelled on the part of a barn owl’s auditory system that uses this time difference. Two incoming signals gave the circuit a question it could answer as the sound arrived. Which horizontal direction best fits the delay between them? The pattern of activity across its outputs represented possible directions. Lazzaro and Mead’s sound-localisation experiment.
Choosing between competing signals was itself something the group could build in silicon. Lazzaro, Sylvie Ryckebusch, Mahowald and Mead tested networks in which the strongest input suppressed weaker responses through a shared wire. The team used these circuits in systems for locating sounds and estimating depth from two views. The engineers chose the rule and could measure how the circuit applied it. For AI, this gave a concrete answer to a design question that still matters in a listening device or a robot. Which part of a perceptual judgement can the electronics make while the signal is arriving, and which part should be left to the next stage? The four researchers’ account of the competing circuits.
Hearing one word, returning one ball
Consider a device that has to listen all day for a single spoken word. In 2024 Hannah Bos and Dylan Muir reported a test of the Xylo Audio 2 processor on the word “Aloha”. Their deployed model achieved 95 per cent accuracy on that test. The extra power drawn by the processor during inference was 291 microwatts. A person buying a battery for the finished device would need to know how much power the whole device draws, including while it waits. The distinction matters most in exactly the places Mead’s approach holds promise, where a sensor has to stay awake for hours. Bos and Muir’s measurements · How NeuroBench compares systems.
Time can matter as much as energy. In a 2026 table-tennis experiment, a robot had an event camera, a tracker and a bat. The camera reported changes in brightness at individual pixels as they occurred. A tracker could locate the ball accurately in a recording and still send its answer too late for the bat to meet it. The researchers measured what happened when perception took longer, right down to whether the robot returned the ball. Their camera used changes over time. The 1988 retina had compared neighbouring points in space. What the two projects share is a decision about which information the sensor should pass on to the rest of the machine. The robot and the timing experiment.
A chronology
- 1934–1960Born in Bakersfield, grows up at Big Creek and earns three electrical engineering degrees at Caltech.
- 1970s–1980Students share fabrication runs. Mead and Conway publish their textbook on integrated circuit design.
- 1988Reports the silicon retina with Misha Mahowald and the electronic cochlea with Richard Lyon.
- 1989–1992Lazzaro and Mead report their sound-localisation chip. Mead publishes his book and his paper on neuromorphic electronics. Mahowald completes her doctoral work on silicon vision.
- 1999–2024Becomes professor emeritus at Caltech in 1999. Receives the 2022 Kyoto Prize for his work on circuit design and a lifetime contribution award for neuromorphic engineering in 2024.
Further reading
For a view from the laboratory bench, read the retina paper alongside Mahowald’s thesis. The first asks how faithfully a single circuit can represent a stage of vision. The second follows the signals as they leave one chip and enter another. In that short journey a model of a cell begins to turn into a machine that can judge distance.
Liked this post? Follow this blog to get more.