In Max and Shen’s study, sensor readings are converted into a stream of electrical spikes. These signals are then sparsely and randomly projected onto a hidden layer that replaces the brain’s Kenyon cells. Instead of relying on a single APL-like control neuron, the model allows neurons in the hidden layer to inhibit one another. The hidden layer then connects to an output layer, with one neuron assigned to each labeled odor. During training, the system learns which neural “barcode” corresponds to each scent.
How Spi-Fly Learns with Simple Rules
The Spi-Fly odor-recognition system learns the relationship between each neural barcode and its label using principles from early artificial neural networks. Whenever a hidden neuron contributes to a correct classification, the connection between that neuron and the corresponding output is strengthened slightly.
That is the entire learning process. Spi-Fly does not require backpropagation, the method used to train many modern neural networks by working backward through every layer to determine which connections contributed to an error. Backpropagation was co-invented by Geoffrey Hinton, who received the 2024 Nobel Prize in Physics.
In testing, the simplified neuromorphic model performed well. When exposed to odors detected by a typical gas sensor, Spi-Fly reached strong classification performance after just three exposures to each scent. A conventional backpropagation-based system required roughly 70 exposures to achieve comparable results.
The model also performed better when learning new odors over time. When researchers introduced unfamiliar scents two at a time—a stress test designed to expose catastrophic forgetting—Spi-Fly retained its earlier odor memories with only a small drop in accuracy. By comparison, the backpropagation system’s performance fell close to random guessing.
Memory capacity is another important consideration. The long-term goal is to run systems such as Spi-Fly on neuromorphic chips, specialized hardware designed to mimic aspects of brain function by processing spikes rather than conventional software instructions. These chips are energy-efficient but typically have limited memory, meaning learning algorithms must operate within tight hardware constraints. Under those conditions, Spi-Fly’s performance degrades far less than that of backpropagation.
However, the system still has a theoretical limit. “If a sparse code layer contains 100 neurons and each odor is represented by five neurons, then the theoretical upper limit for odor capacity is ‘100, select 5,’” Max explained in an email. That creates approximately 75 million possible neural “barcodes.” In practice, however, environmental noise eliminates much of that capacity. Airflow changes and sensor variability mean the same odor may produce a different sparse code each time it is detected. “Given the parameters of our model, we should be able to represent hundreds of odors,” Max wrote, “but it also depends on the type of odor and the similarity of the odors.”
Source: arstechnica.com