Lazy learning: a biologically-inspired plasticity rule for fast and energy efficient synaptic plasticity

When training neural networks for classification tasks with backpropagation, parameters are updated on every trial, even if the sample is classified correctly. In contrast, humans concentrate their learning effort on errors. Inspired by human learning, we introduce lazy learning, which only learns on incorrect samples. Lazy learning can be implemented in a few lines of code and requires no hy-perparameter tuning. Lazy learning achieves state-of-the-art performance and is particularly suited when datasets are large. For instance, it reaches 99.2% test accuracy on Extended MNIST using a single-layer MLP, and does so 7.6× faster than a matched backprop network. Recent progress in machine learning has been partly attributed to the use of large data sets [LeCun et al., 2015]. Even already large datasets are often augmented to further boost performance. However, repeatedly cycling over large datasets and adjusting the parameters is time and energy consuming. In classification tasks, backprop typically prescribes synaptic updates regardless of whether the classification was correct or incorrect; updating the network to be correct if it was wrong, but also updating to be more correct if it was right. Is this incessant updating necessary or do more efficient training rules exist? Previous work has determined that high accuracies for many datasets can be achieved using only a small number of critical samples, called the coreset [Agarwal et al., 2005]. Methods have been developed to find these samples and focus training [Toneva et al., 2019, Paul et al., 2021, Coleman et al., 2020]. These methods require a model trained beforehand to identify key samples and form a coreset dataset. However, for a new dataset, it won't be clear what the coreset is or even whether the coreset is small or large. A simple algorithm that restricts learning and does not need tuning is lacking. Humans appear to be more efficient by placing far more importance on mistakes and errors, a phenomenon known as the negativity bias [Rozin and Royzman, 2001]. A familiar example is the sharp 'jolt' we experience when typing a word incorrectly. When such mistakes occur a large event-related potential known as the error-related negativity (ERN) is elicited in the electroencephalogram signal (EEG), which usually precedes behavioural changes like post-error slowing and post-error improvement in accuracy [Kalfao˘ glu et al., 2017]. The ERN is linked to dopamine [Holroyd and Coles, 2002]-a plasticity modulator required for persistent forms of plasticity [O'Carroll and Morris, 2004]. Indeed, subjects learn better on stimuli that evoke a larger EEG [de Bruijn et al., 2020]. This suggests that, in contrast to backprop, biological learning largely occurs on mistakes, while skipping over correct samples. The reason might be that learning is emerging as a metabolically costly process. For instance, flies subjected to aversive conditioning and subsequently starved, died 20% sooner