Original Learning Task Research Articles

Efficient Vertical Federated Unlearning via Fast Retraining

Vertical federated learning (VFL) revolutionizes privacy-preserved collaboration for small businesses, that have distinct but complementary feature sets. However, as the scope of VFL expands, the constant entering and leaving of participants, as well as the subsequent exercise of the “right to be forgotten” pose a great challenge in practice. The question of how to efficiently erase one’s contribution from the shared model remains largely unexplored in the context of vertical federated learning. In this paper, we introduce a vertical federated unlearning framework, which integrates model checkpointing techniques with a hybrid, first-order optimization technique. The core concept is to reduce backpropagation time and improve convergence/generalization by combining the advantages of the existing optimizers. We provide in-depth theoretical analysis and time complexity to illustrate the effectiveness of the proposed design. We conduct extensive experiments on 6 public datasets and demonstrate that our method could achieve up to 6.3 × speed-up compared to the baseline, with negligible influence on the original learning task.

Reinforcing Neural Network Stability with Attractor Dynamics

Recent approaches interpret deep neural works (DNNs) as dynamical systems, drawing the connection between stability in forward propagation and generalization of DNNs. In this paper, we take a step further to be the first to reinforce this stability of DNNs without changing their original structure and verify the impact of the reinforced stability on the network representation from various aspects. More specifically, we reinforce stability by modeling attractor dynamics of a DNN and propose relu-max attractor network (RMAN), a light-weight module readily to be deployed on state-of-the-art ResNet-like networks. RMAN is only needed during training so as to modify a ResNet's attractor dynamics by minimizing an energy function together with the loss of the original learning task. Through intensive experiments, we show that RMAN-modified attractor dynamics bring a more structured representation space to ResNet and its variants, and more importantly improve the generalization ability of ResNet-like networks in supervised tasks due to reinforced stability.

The interplay between self-regulation in learning and cognitive load

Research on self-regulated learning and on cognitive load has been two of the most prominent and influential research lines in educational research during the last decades. However, both lines developed quite independently from one another. This paper aims to bridge both concepts in order to better understand self-regulation as well as cognitive affordances of complex and dynamic learning processes. In fact, most learning environments require learners to self-regulate their learning process. They have to set their goals and plan, use strategies and monitor their learning progress and if necessary they have to regulate, i.e. to adapt their learning behavior. However, in all these phases of self-regulation, learners need to invest cognitive and metacognitive resources in addition to dealing with the original learning task. The affordances of self-regulation thus will cause cognitive load. This paper analyzes the affordances in the different phases of self-regulated learning in terms of intrinsic, extraneous and germane load. As a conclusion, the interplay between different affordances of self-regulation and learners’ resources and aptitudes is described in an integrated model of self-regulation and cognitive load.

Controlled permutations for testing adaptive learning models

We study evaluation of supervised learning models that adapt to changing data distribution over time (concept drift). The standard testing procedure that simulates online arrival of data (test-then-train) may not be sufficient to generalize about the performance, since that single test concludes how well a model adapts to this fixed configuration of changes, while the ultimate goal is to assess the adaptation to changes that happen unexpectedly. We propose a methodology for obtaining datasets for multiple tests by permuting the order of the original data. A random permutation is not suitable, as it makes the data distribution uniform over time and destroys the adaptive learning task. Therefore, we propose three controlled permutation techniques that make it possible to acquire new datasets by introducing restricted variations in the order of examples. The control mechanisms with theoretical guarantees of preserving distributions ensure that the new sets represent close variations of the original learning task. Complementary tests on such sets allow to analyze sensitivity of the performance to variations in how changes happen and this way enrich the assessment of adaptive supervised learning models.

Remembered Utility of Unpleasant and Pleasant Learning Experiences: Is All Well that Ends Well?

SummaryRetrospective evaluation of the pain or pleasure associated with past episodes (i.e. remembered utility) is mainly determined by how the episodes felt when they were at their peak, and when they ended (Kahneman, Fredrickson, Schreiber, & Redelmeier, 1993). Two experiments were conducted to investigate whether this memory bias known as the peak‐end rule also affects the remembered utility of learning episodes and shapes study behaviour. Both the first experiment, which focused on unpleasant learning experiences imposing a very high cognitive load, and the second experiment, which focused on pleasant learning experiences imposing a low cognitive load, presented primary school children with a short study list and an extended study list with English–Dutch word pairs. Results showed that study experiences that ended on an experienced peak affect were evaluated as easier to learn and easier to cope with. These measures of remembered utility shaped children's immediate and delayed study behaviour. In addition, the pattern of affect induced by the word pairs that were added to the extended study list in both experiments suggests that the experience with new learning tasks is relative to the affect induced by the original learning task. The findings are congruent with the peak‐end rule and support the position that the design of learning environments could be improved by taking remembered utility into account. Copyright © 2012 John Wiley & Sons, Ltd.

WHATEVER GAVE YOU THAT IDEA? FALSE MEMORIES FOLLOWING EQUIVALENCE TRAINING: A BEHAVIORAL ACCOUNT OF THE MISINFORMATION EFFECT

The misinformation effect is a term used in the cognitive psychological literature to describe both experimental and real-world instances in which misleading information is incorporated into an account of an historical event. In many real-world situations, it is not possible to identify a distinct source of misinformation, and it appears that the witness may have inferred a false memory by integrating information from a variety of sources. In a stimulus equivalence task, a small number of trained relations between some members of a class of arbitrary stimuli result in a large number of untrained, or emergent relations, between all members of the class. Misleading information was introduced into a simple memory task between a learning phase and a recognition test by means of a match-to-sample stimulus equivalence task that included both stimuli from the original learning task and novel stimuli. At the recognition test, participants given equivalence training were more likely to misidentify patterns than those who were not given such training. The misinformation effect was distinct from the effects of prior stimulus exposure, or partial stimulus control. In summary, stimulus equivalence processes may underlie some real-world manifestations of the misinformation effect.

Memory consolidation and reconsolidation of an inhibitory avoidance task in mice: Effects of a new different learning task

CF-1 male mice were trained in an inhibitory avoidance task using a high footshock (1.2mA, 50Hz, 1 s) in order to reduce the influence of extinction on retention performance. A single session of 5 min exposure to a hole-board (nose-poke behavior), either immediately after training or the first retention test (memory reactivation) impaired retention performance over two consecutive days. The effects were time-dependent since they were not observed when the exposure to the hole-board was delayed 3 h. When mice were habituated to the hole-board (5 min/day, 5 days), and then trained in an inhibitory avoidance task, the immediately post-training or memory reactivation exposure to the hole-board did not modify retention performance over two consecutive days. The effects of the post-reactivation acute exposure to the hole-board were long-lasting (21 days). Reinstatement was not observed in our experimental conditions. The non-spontaneous recovery of retention performance over 21-days and the lack of reinstatement, suggest that the impairment of retention performance observed was not probably due to a deficit in memory retrieval. These findings suggest that the exposure to a potential new learning situation impairs not only memory consolidation but also memory reconsolidation of the original learning task.

Position reversal learning in aged Japanese macaques

We examined aged and young monkeys using a multiple position reversal task to investigate declines in cognitive functions with aging. The task consisted of an original learning task (simple position discrimination task) and a reversal learning task. While the performance of the aged monkeys was not different from that of the young monkeys in the original learning task, the aged monkeys showed a poorer performance than the young monkeys in the reversal learning task. According to our response analysis, the poor performance of aged monkeys in the reversal learning was not caused mainly by repetition of error responses, but rather by the impairment of understanding of the association between stimulus and reward. These results suggest that the prefrontal cortex, particularly the medial orbital cortex, is impaired with aging.

Discrimination behavior of rats with septal lesions in a cue addition deletion paradigm

The discrimination behavior of rats with septal lesions is affected by their altered reactivity to stimulus conditions surrounding the learning situation. Compared to control animals rats with septal lesions are differentially affected by experience with cues relevant to the task itself. In the present experiment we examined the effects of septal lesions on how the rat responds to, and subsequently utilizes, cue information which is made relevant after the animal had achieved criterion performance in a discrimination task. Relative to control animals, rats with septal lesions apparently failed to attend to the newly relevant cues provided, and clearly were deficient when later called upon to utilize such information. However, the degree of deficit was related to the difficulty exhibited in acquisition of original learning. Thus, the more difficult the original learning task, the less deficient were lesioned animals compared to their control counterparts in subsequent situations demanding the utilization of new information. These findings are consistent with our notion that septal lesions alter the relative weighting placed on specific information which the CNS must process. Changes in learning performance reflect altered sensory reactivity.

Age effects in the acquisition and retention of active and passive avoidance learning by rats

AbstractThree groups of 8 male albino rats within each of the age ranges 24, 50, and 100 days were trained on an active avoidance task to a criterion of 10 successive avoidances. The number of trials to criterion was found to be a monotonic function of age. One group at each age was tested immediately thereafter for passive avoidance of the conditioning chamber, another group was tested for passive avoidance after 25 days, and the third was tested for retention of active avoidance after 25 days. Groups at 50, 75, and 125 days were used as controls for the active avoidance retention groups. All animals took significantly fewer trials to relearn the active avoidance after 25 days than did controls on the acquisition of the original learning task. On passive avoidance 24‐day olds performed less efficiently than the older groups and failed to show any retention after 25 days.

Strength of dimensional preferences as a predictor of nursery-school children's performance on a concept-shift task

The strength of nursery-school children's dimensional preferences for form and color was scaled by a multidimensional technique, revealing a wide range of strength of preference. Unlike earlier studies, many Ss were found to have equivalent preference for both dimensions. All Ss were then given a two-choice simultaneous discrimination learning problem with either form or color relevant to solution followed by an intradimensional (ID) or an extradimensional (ED) shift. In general, the effects of scaled preferences were in the same direction as, but stronger than, the effects of preference reported in previous studies. The performance of the equal-preference Ss on the original learning task was equivalent to that of Ss who strongly preferred the relevant dimension, and equal-preference Ss learned ID and ED shifts equally well, suggesting that equal-preference Ss may be more flexible in their problem-solving than Ss with a preference for a single dimension. It was further suggested that an age-related hierarchy of preferences progresses from color- to form- to equal-preference.