A contrastive rule for meta-learning

Abstract

Meta-learning algorithms leverage regularities that are present on a set of tasks to speed up and improve the performance of a subsidiary learning process. Recent work on deep neural networks has shown that prior gradient-based learning of meta-parameters can greatly improve the efficiency of subsequent learning. Here, we present a biologically plausible meta-learning algorithm based on equilibrium propagation. Instead of explicitly differentiating the learning process, our contrastive meta-learning rule estimates meta-parameter gradients by executing the subsidiary process more than once. This avoids reversing the learning dynamics in time and computing second-order derivatives. In spite of this, and unlike previous first-order methods, our rule recovers an arbitrarily accurate meta-parameter update given enough compute. We establish theoretical bounds on its performance and present experiments on a set of standard benchmarks and neural network architectures.

The least-control principle for local learning at equilibrium

“…We apply our principle to learn equilibrium recurrent neural networks with a variety of architectures, from laterally-connected networks of leaky integrator neurons to deep equilibrium models [8][9][10], a recent family of high-performance models which repeat until convergence a sequence of more complex computations. To further demonstrate the generality of our principle, we then consider a recently studied meta-learning problem where the goal is to change the internal state of a complex synapse such that future learning performance is improved [30]. We find that our single-phase local learning rules yield highly competitive performance in both application domains.…”

“…A seminal result shows that as β → 0 the gradient ∇ θ L(φ * ) of the original objective function is recovered [24,39]. However, this update is known to be sensitive to noise as the value of f can be very small at the weakly-nudged equilibrium [30]. Our least-control principle works at the opposite perfect control end of the spectrum (β → ∞), and it is therefore more resistant to noise.…”

“…Here, we show that it can be applied to metalearning, a framework that aims at improving a learning algorithm by leveraging shared structure in encountered tasks [63]. This framework is gaining traction in neuroscience, having been featured in recent work aiming at understanding which systems [64][65][66] and rules [30] could support meta-learning in the brain.…”

Predicting Flat-Fading Channels via Meta-Learned Closed-Form Linear Filters and Equilibrium Propagation

“…Specifically, we first develop a low-complexity offline solution that obtains hyperparameters and predictors in closed form. Then, an online algorithm is proposed based on gradient descent and equilibrium propagation (EP) [7], [8]. Previous applications of meta-learning to communication systems include demodulation [9], [10], channel equalization [11], encoding/decoding [12], [13], [14], MIMO detection [15], beamforming [16], [17], and resource allocation [18].…”

Bottom-Up and Top-Down Neural Processing Systems Design: Neuromorphic Intelligence as the Convergence of Natural and Artificial Intelligence

“…This growth rate was much faster than that of technology scaling, and outweighed the efforts to reduce the network computational footprint [16]. In order to improve the ability of ANN-based AI to scale, diversify, and generalize from limited data while avoiding catastrophic forgetting, meta-learning approaches are investigated [17]- [22]. These approaches aim at building systems that are tailored to their environment and can quickly adapt once deployed, just as evolution shapes the degrees of versatility and online adaptation of biological brains [23].…”