An Analog VLSI Model of the Fly Elementary Motion Detector

Reid R. Harrison, Christof KochNeurIPS 1997

TL;DR

Researchers have successfully implemented a biological "Elementary Motion Detector" (EMD) on an analog VLSI chip. By mimicking the neural architecture of a fly’s eye, they created a low-power, real-time sensor that reproduces the spatial, temporal, and directional tuning seen in nature.

What Problem Does This Paper Address?

In the world of robotics and autonomous systems, detecting motion quickly and efficiently is a massive challenge. Traditional computer vision often relies on power-hungry digital processors to crunch through frames of video. In contrast, a common housefly performs complex maneuvers—like landing on a moving object or stabilizing its flight—using a tiny, low-power nervous system that is exceptionally robust.

The "gap" the authors address is the lack of efficient hardware building blocks for these tasks. They aim to bridge the divide between biological models of vision (specifically the Reichardt correlation model) and practical engineering by building a physical, analog silicon circuit that "thinks" like a fly.

Methods

The authors implemented an elaborated version of the Reichardt motion detector using a 2.0-$\mu$m analog CMOS VLSI process. The architecture follows a specific pipeline:

  1. Integrated Photoreceptors: Adaptive circuits that produce a voltage proportional to the log of light intensity.
  2. High-Pass Filtering: A temporal derivative circuit that removes DC offsets (static backgrounds) and splits the signal into "ON" and "OFF" channels, mimicking biological visual pathways.
  3. Low-Pass Filtering (The Delay): To detect motion, you need to compare a signal at one point in space with a delayed signal from a neighboring point. The authors used the inherent phase lag of a current-mode low-pass filter to provide this "delay."
  4. Multiplication & Subtraction: Translinear circuits (using transistors in their subthreshold region) multiply the signals from adjacent receptors. Subtracting the outputs of two mirror-symmetric correlators allows the chip to distinguish direction and ignore full-field flicker.

A notable technical detail is the use of floating-gate nFETs in the multipliers, which allows for precise current-mode computation by equalizing charges via UV light exposure.

Datasets

As this is a paper about hardware synthesis rather than machine learning, there are no "training" datasets. Instead, the authors evaluated their chip using:

  • Synthetic Stimuli: Drifting sinusoidal gratings printed on a rotating drum to test frequency and directional responses.
  • Biological Benchmarks: The chip's performance was compared against existing electrophysiological data from the wide-field motion neurons (H1 and HS cells) of the blowfly (Calliphora erythrocephala) and the hoverfly (Volucella pelluscens).

Key Insights and Ideas

  • Phase Lag as Delay: One of the "aha" moments is using a simple low-pass filter's phase lag to substitute for the complex neural delays in biological systems. It’s a computationally "cheap" way to implement a Reichardt detector in silicon.
  • Current-Mode Advantages: By switching from voltage-mode to current-mode representation, the authors utilized translinear circuits. These are naturally suited for multiplication and power-law computations while avoiding the fixed offset errors common in other analog designs.
  • Bio-Realistic Aliasing: The chip exhibits spatial aliasing (responding in the wrong direction at very high spatial frequencies), which is a known phenomenon in insect vision. This suggests the silicon model is capturing the fundamental constraints of the biological system.

Key Results

  • Temporal Tuning: The circuit's peak response is determined by the low-pass filter's bias current, allowing the "speed" of the detector to be tuned electronically.
  • Directional Selectivity: The chip’s response to motion direction follows a near-perfect cosine function, matching the directional tuning of the fly’s H1 neuron.
  • Real-Time Performance: The system operates in continuous time with negligible power consumption, making it a viable building block for mobile, battery-operated robots.
  • Qualitative Match: The spatial and temporal frequency response curves of the silicon chip closely mirror those of real blowflies, validating the neuromorphic approach.

Limitations and Future Work

The authors identify a few hurdles before these chips can be deployed in large arrays:

  • Gain Mismatch: There is currently a 2.5x gain mismatch between the left and right channels that requires manual compensation. Future iterations will need better transistor matching or auto-calibration.
  • Contrast Saturation: Unlike real flies, which show a saturating response to high-contrast patterns, this circuit remains linear. The authors suggest adding saturating nonlinearities in future versions to better mimic biological robustness.
  • Scaling: While a single EMD works well, the next step is integrating hundreds of these into a retinotopic array to perform "spatial integration," which is how flies achieve such high sensitivity to motion.

Summary updated

BibTeX
@inproceedings{harrison1997analog,
  author = {Reid R. Harrison and Christof Koch},
  title = {{An Analog VLSI Model of the Fly Elementary Motion Detector}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1997},
}
Figure