The power of feedback.

A major open problem on the road to artificial intelligence is the development of incrementally learning systems that learn about more and more concepts over time from a stream of data. In this work, we introduce a new training strategy, iCaRL, that allows learning in such a class-incremental way: only the training data for a small number of classes has to be present at the same time and new classes can be added progressively. iCaRL learns strong classifiers and a data representation simultaneously. This distinguishes it from earlier works that were fundamentally limited to fixed data representations and therefore incompatible with deep learning architectures. We show by experiments on CIFAR-100 and ImageNet ILSVRC 2012 data that iCaRL can learn many classes incrementally over a long period of time where other strategies quickly fail.

/pdf/icarl-incremental-classifier-and-representation-learning-4iat42kiu2.pdf

iCaRL: Incremental Classifier and Representation Learning

Continual lifelong learning with neural networks: A review.

Over more than a century of research has established the fact that sleep benefits the retention of memory. In this review we aim to comprehensively cover the field of "sleep and memory" research by providing a historical perspective on concepts and a discussion of more recent key findings. Whereas initial theories posed a passive role for sleep enhancing memories by protecting them from interfering stimuli, current theories highlight an active role for sleep in which memories undergo a process of system consolidation during sleep. Whereas older research concentrated on the role of rapid-eye-movement (REM) sleep, recent work has revealed the importance of slow-wave sleep (SWS) for memory consolidation and also enlightened some of the underlying electrophysiological, neurochemical, and genetic mechanisms, as well as developmental aspects in these processes. Specifically, newer findings characterize sleep as a brain state optimizing memory consolidation, in opposition to the waking brain being optimized for encoding of memories. Consolidation originates from reactivation of recently encoded neuronal memory representations, which occur during SWS and transform respective representations for integration into long-term memory. Ensuing REM sleep may stabilize transformed memories. While elaborated with respect to hippocampus-dependent memories, the concept of an active redistribution of memory representations from networks serving as temporary store into long-term stores might hold also for non-hippocampus-dependent memory, and even for nonneuronal, i.e., immunological memories, giving rise to the idea that the offline consolidation of memory during sleep represents a principle of long-term memory formation established in quite different physiological systems.

https://www.zora.uzh.ch/id/eprint/77770/1/physiol_rev_93.pdf

About sleep's role in memory

http://www.bcp.psych.ualberta.ca/%7Emike/Pearl_Street/PSYCO354/pdfstuff/Readings/French2.pdf

Catastrophic forgetting in connectionist networks.

In this paper we review the literature relating to the psychological/educational study of programming. We identify general trends comparing novice and expert programmers, programming knowledge and strategies, program generation and comprehension, and object-oriented versus procedural programming. (We do not cover research relating specifically to other programming styles.) The main focus of the review is on novice programming and topics relating to novice teaching and learning. Various problems experienced by novices are identified, including issues relating to basic program design, to algorithmic complexity in certain language features, to the “fragility” of novice knowledge, and so on. We summarise this material and suggest some practical implications for teachers. We suggest that a key issue that emerges is the distinction between effective and ineffective novices. What characterises effective novices? Is it possible to identify the specific deficits of ineffective novices and help them to become effec...

/pdf/learning-and-teaching-programming-a-review-and-discussion-2x8svsn4u8.pdf

Learning and Teaching Programming: A Review and Discussion

This paper reviews the problem of catastrophic forgetting (the loss or disruption of previously learned information when new information is learned) in neural networks, and explores rehearsal mechanisms (the retraining of some of the previously learned information as the new information is added) as a potential solution. We replicate some of the experiments described by Ratcliff (1990), including those relating to a simple 'recency' based rehearsal regime. We then develop further rehearsal regimes which are more effective than recency rehearsal. In particular, 'sweep rehearsal' is very successful at minimizing catastrophic forgetting. One possible limitation of rehearsal in general, however, is that previously learned information may not be available for retraining. We describe a solution to this problem, 'pseudorehearsal', a method which provides the advantages of rehearsal without actually requiring any access to the previously learned information (the original training population) itself. We then sugge...

/pdf/catastrophic-forgetting-rehearsal-and-pseudorehearsal-4e9ku43sdc.pdf

Catastrophic forgetting, rehearsal and pseudorehearsal

http://www.cs.otago.ac.nz/staffpriv/anthony/publications/pdfs/AbrahamRobins.pdf

Memory retention – the synaptic stability versus plasticity dilemma

Compared to other subjects, the typical introductory programming (CS1) course has higher than usual rates of both failing and high grades, creating a characteristic bimodal grade distribution. In this article, I explore two possible explanations. The conventional explanation has been that learners naturally fall into populations of programmers and non-programmers. A review of decades of research, however, finds little or no evidence to support this account. I propose an alternative explanation, the learning edge momentum (LEM) effect. This hypothesis is introduced by way of a simulated model of grade distributions, and then grounded in the psychological and educational literature. LEM operates such that success in acquiring one concept makes learning other closely linked concepts easier (whereas failure makes it harder). This interaction between the way that people learn and the tightly integrated nature of the concepts comprising a programming language creates an inherent structural bias in CS1, which dr...

/pdf/learning-edge-momentum-a-new-account-of-outcomes-in-cs1-xu6oeqijl0.pdf

Learning edge momentum: a new account of outcomes in CS1

In this paper we describe an ongoing study of novice programmers. The aim is to record (as close as possible to) all of the problems encountered by students during the laboratory sessions of our introductory Java programming class. We discuss the tools and methods employed, in particular presenting the list of problem definitions which is used to classify students' problems. Data collected during 2003 are presented and discussed. The results are consistent with trends noted in the literature, and highlight the significance of both fundamental design issues and the procedural aspects of programming. Different problem distributions are observed for high and low performing students. An analysis of individual lab sessions can be useful for refining course materials and teaching practice.

/pdf/my-program-is-correct-but-it-doesn-t-run-a-preliminary-401efo7ap1.pdf

Anthony Robins

Papers

Learning and Teaching Programming: A Review and Discussion

Catastrophic forgetting, rehearsal and pseudorehearsal

Memory retention – the synaptic stability versus plasticity dilemma

Learning edge momentum: a new account of outcomes in CS1

My program is correct but it doesn't run: a preliminary investigation of novice programmers' problems