Modelling continual learning in humans with Hebbian context gating and exponentially decaying task signals.

Humans can learn several tasks in succession with minimal mutual interference but perform more poorly when trained on multiple tasks at once. The opposite is true for standard deep neural networks. Here, we propose novel computational constraints for artificial neural networks, inspired by earlier work on gating in the primate prefrontal cortex, that capture the cost of interleaved training and allow the network to learn two tasks in sequence without forgetting. We augment standard stochastic gradient descent with two algorithmic motifs, so-called "sluggish" task units and a Hebbian training step that strengthens connections between task units and hidden units that encode task-relevant information. We found that the "sluggish" units introduce a switch-cost during training, which biases representations under interleaved training towards a joint representation that ignores the contextual cue, while the Hebbian step promotes the formation of a gating scheme from task units to the hidden layer that produces orthogonal representations which are perfectly guarded against interference. Validating the model on previously published human behavioural data revealed that it matches performance of participants who had been trained on blocked or interleaved curricula, and that these performance differences were driven by misestimation of the true category boundary.

Tracking the contribution of inductive bias to individualised internal models.

Internal models capture the regularities of the environment and are central to understanding how humans adapt to environmental statistics. In general, the correct internal model is unknown to observers, instead they rely on an approximate model that is continually adapted throughout learning. However, experimenters assume an ideal observer model, which captures stimulus structure but ignores the diverging hypotheses that humans form during learning. We combine non-parametric Bayesian methods and probabilistic programming to infer rich and dynamic individualised internal models from response times. We demonstrate that the approach is capable of characterizing the discrepancy between the internal model maintained by individuals and the ideal observer model and to track the evolution of the contribution of the ideal observer model to the internal model throughout training. In particular, in an implicit visuomotor sequence learning task the identified discrepancy revealed an inductive bias that was consistent across individuals but varied in strength and persistence.

Optimal forgetting: Semantic compression of episodic memories.

It has extensively been documented that human memory exhibits a wide range of systematic distortions, which have been associated with resource constraints. Resource constraints on memory can be formalised in the normative framework of lossy compression, however traditional lossy compression algorithms result in qualitatively different distortions to those found in experiments with humans. We argue that the form of distortions is characteristic of relying on a generative model adapted to the environment for compression. We show that this semantic compression framework can provide a unifying explanation of a wide variety of memory phenomena. We harness recent advances in learning deep generative models, that yield powerful tools to approximate generative models of complex data. We use three datasets, chess games, natural text, and hand-drawn sketches, to demonstrate the effects of semantic compression on memory performance. Our model accounts for memory distortions related to domain expertise, gist-based distortions, contextual effects, and delayed recall.

Measuring and filtering reactive inhibition is essential for assessing serial decision making and learning.

Learning complex structures from stimuli requires extended exposure and often repeated observation of the same stimuli. Learning induces stimulus-dependent changes in specific performance measures. The same performance measures, however, can also be affected by processes that arise because of extended training (e.g., fatigue) but are otherwise independent from learning. Thus, a thorough assessment of the properties of learning can only be achieved by identifying and accounting for the effects of such processes. Reactive inhibition is a process that modulates behavioral performance measures on a wide range of time scales and often has opposite effects than learning. Here we develop a tool to disentangle the effects of reactive inhibition from learning in the context of an implicit learning task, the alternating serial reaction time (ASRT) task. Our method highlights that the magnitude of the effect of reactive inhibition on measured performance is larger than that of the acquisition of statistical structure from stimuli. We show that the effect of reactive inhibition can be identified not only in population measures but also at the level of performance of individuals, revealing varying degrees of contribution of reactive inhibition. Finally, we demonstrate that a higher proportion of behavioral variance can be explained by learning once the effects of reactive inhibition are eliminated. These results demonstrate that reactive inhibition has a fundamental effect on the behavioral performance that can be identified in individual participants and can be separated from other cognitive processes like learning. (PsycINFO Database Record