Algorithmic insights on continual learning from fruit flies

Abstract

Continual learning in computational systems is challenging due to catastrophic forgetting. We discovered a two-layer neural circuit in the fruit fly olfactory system that addresses this challenge by uniquely combining sparse coding and associative learning. In the first layer, odors are encoded using sparse, high-dimensional representations, which reduces memory interference by activating non-overlapping populations of neurons for different odors. In the second layer, only the synapses between odor-activated neurons and the odor's associated output neuron are modified during learning; the rest of the weights are frozen to prevent unrelated memories from being overwritten. We show empirically and analytically that this simple and light-weight algorithm significantly boosts continual learning performance. The fly's associative learning algorithm is strikingly similar to the classic perceptron learning algorithm, albeit two modifications, which we show are critical for reducing catastrophic forgetting. Overall, fruit flies evolved an efficient life-long learning algorithm, and circuit mechanisms from neuroscience can be translated to improve machine computation.

Informing generative replay for continual learning with long-term memory formation in the fruit fly

“…To investigate performance of the generative replay model, we extend the modeling and continual learning investigation approach utilized for the studying the feedforward portion of this model [12]. A class-incremental learning CIFAR-100 baseline task (25 training experiences of 4 classes each) is used where processed sensory input is Imagenet pre-trained Resnet embeddings (dimensionality 512), as further detailed in [12]). A 10% sparse binary matrix is used for W xh and the nonlinearity g(.)…”

“…MBON activation drives higher-level behavioral decisions and associative learning at the synapses between KC's and MBON's is a substrate for learning to link sensory input patterns with rewarded behavioral responses. From a continual learning perspective, there has been recent work which identified algorithmic insights in the feedforward path of this system [12]. The full connectivity of the mushroom body is complex though, with specific recurrent connections of protocerebral anterior medial (PAM) neurons in the α1 compartment having been identified to be necessary for long-term memory (LTM) formation [5].…”

Exploiting large neuroimaging datasets to create connectome-constrained approaches for more robust, efficient, and adaptable artificial intelligence

“…Towards this end, our team has analyzed circuitry for sensory memory formation in the Hemibrain dataset, building on observations of feedback reward circuitry in the fruit fly 44 and prior modeling work on feedforward circuitry. 45 Utilizing novel feedback connectivity observed at the neuron-synapse level in the Hemibrain dataset, the team designed a novel generative replay approach utilizing feedback connections. This replay approach resulted in an over 20% accuracy improvement in an incremental task learning scenario with the CIFAR-100 dataset, approaching the performance of upper-bound baselines.…”

A Theoretical Perspective on Hyperdimensional Computing

“…Neurally-inspired lightweight algorithms have recently been proposed for lifelong learning applications. FlyModel [52] and SDMLP [8] use sparse coding and associative memory for lifelong learning. However, both approaches assume full supervision.…”

“…In HDC, all data is represented using high-dimensional, low-precision (often binary) vectors known as "hypervectors, " which can be manipulated through simple elementwise operations to perform tasks like memorization and learning. HDC is well-understood from a theoretical standpoint [56] and shares intriguing connections with biological lifelong learning [52]. Furthermore, its use of basic element-wise operators aligns with highly parallel and energy-efficient hardware, offering substantial energy savings in IoT applications [11,23,27,65].…”