DynMat, a network that can learn after learning

Abstract

To survive in the dynamically-evolving world, we accumulate knowledge and improve our skills based on experience. In the process, gaining new knowledge does not disrupt our vigilance to external stimuli. In other words, our learning process is 'accumulative' and 'online' without interruption. However, despite the recent success, artificial neural networks (ANNs) must be trained offline and suffer catastrophic interference between old and new learning, indicating that ANNs' conventional learning algorithms may not be suitable for building intelligent agents comparable to our brain. In this study, we propose a novel neural network architecture (DynMat) consisting of dual learning systems inspired by the complementary learning system (CLS) theory suggesting that the brain relies on short-and long-term learning systems to learn continuously.Our empirical evaluations show that 1) DynMat can learn a new class without catastrophic interference and 2) it does not strictly require offline training.building continuous learning systems remains difficult. Moreover, catastrophic interference and a few-shot (i.e., rapid) learning have been studied separately.Then, how does the brain learn continuously? The complementary learning system (CLS) theory proposes that the brain utilizes short-and long-term learning systems for continuous (or continual/lifelong) learning (Norman & O'Reilly, 2003;O'Reilly, Bhattacharyya, Howard, & Ketz, 2014). The short-term learning system relies on fast memory to store new experience (or examples) using sparse and non-overlapping codes (representations); see also (Parisi et al., 2018).Due to its sparse representations, the interference between old and new items (i.e., knowledge) is minimized. The information stored in the short-term learning system can be replayed into a more effective learning system which utilizes dense and overlapping codes for information storage.

Neural network process simulations support a distributed memory system and aid design of a novel computer adaptive digital memory test for preclinical and prodromal Alzheimer’s disease.

“…To our knowledge, the use of cascaded architectures as described by Pollack (1987) is not common in modern AI approaches and mirroring the inputs (duplicating the inputs to both the output networks and context networks), as done in our MCN models, is a novel approach. The MCN architecture, when combined with the addition of context inputs, provides a new alternative for solving the common problem of catastrophic interference in neural networks (Beer & Barak, 2019; Ellefsen et al, 2015; Lee, 2019). Future work is needed to explore how mirrored cascaded architectures can be applied to increase efficiency and consistency of modern AI data fitting.…”

Library network, a possible path to explainable neural networks

“…That is, we need effective methods to estimate similarities among HAPs. In doing so, we turn to our earlier short-term memory systems [16].…”

“…Our short-term memory systems [16] store novel input patterns (i.e., substantially different from stored input patterns), and their outputs reflect cosine similarities between a present input and stored inputs. If HAPs, evoked by the same class input patterns, are well clustered together, the stored input patterns in the memory systems can approximate input patterns' (i.e., HAPs) distribution.…”