Scientists at the Okinawa Institute of Science and Technology (OIST) have introduced a novel algorithm, Spi-Fly, designed to overcome the problem of catastrophic forgetting in artificial intelligence. This phenomenon, also known as catastrophic interference, causes neural networks to abruptly lose previously learned information when acquiring new tasks. The Spi-Fly algorithm draws inspiration from the highly efficient olfactory system of the fruit fly, Drosophila melanogaster, to achieve continuous learning with reduced memory interference.
Catastrophic forgetting poses a significant hurdle for AI systems that need to adapt to dynamic, real-world environments, such as autonomous vehicles and large language models. When an AI model is sequentially trained on new data, the optimization process often overwrites the mathematical representations crucial for earlier tasks, leading to performance degradation. For instance, a model initially trained to recognize pedestrians might forget this ability after being fine-tuned to identify new traffic signs.
The fruit fly's brain, despite its compact size of approximately 140,000 neurons, demonstrates remarkable efficiency in processing sensory information and forming associative memories. The OIST researchers focused on the fruit fly's olfactory system, which employs a computational scheme called sparse coding. In this biological mechanism, when a fruit fly encounters an odor, sensory neurons send signals that are then processed such that only a small, specific subset of neurons remains active for that particular scent. This creates a unique "neural barcode" for each smell, minimizing overlap and interference between different odor representations.
Spi-Fly replicates this sparse coding strategy within a simplified three-layer neural network. Incoming data is randomly projected onto a hidden layer containing 1,000 artificial neurons. For each input, only a few of these neurons become active, generating a distinct pattern that the system learns to recognize. This sparse activation helps separate different inputs and maintain low energy consumption, mirroring the fruit fly's biological efficiency.
The algorithm also incorporates a mechanism inspired by the fruit fly's associative learning. In the second layer of the fruit fly's olfactory circuit, only the synapses connecting odor-activated neurons to the output neuron associated with that odor are modified during learning. The remaining synaptic weights are frozen, preventing new learning from overwriting unrelated, previously stored memories. This partial freezing of weights is a key component in Spi-Fly's ability to retain old information while learning new tasks.
The OIST team, including co-author Dr. Yang Shen, published their findings in Neuromorphic Computing & Engineering. Their research indicates that Spi-Fly outperforms several existing class-incremental continual learning algorithms on benchmark datasets like MNIST-20 and CIFAR-100. Notably, the algorithm achieves this without relying on external memory, generative replay, or backpropagation, techniques commonly used to mitigate catastrophic forgetting in other AI models. This makes Spi-Fly a lightweight and computationally efficient solution.
The effectiveness of Spi-Fly was particularly evident in "few-shot learning" scenarios, where the algorithm successfully identified scents after being exposed to only a limited number of examples. This capability is significant for real-world applications where extensive training data may not always be available. The development of Spi-Fly could contribute to the creation of artificial noses capable of classifying a wide range of scents, from detecting spoiled food to identifying hazardous substances, under realistic constraints of limited training data and energy.