Don't Look Back in Anger: MAGIC Net for Streaming Continual Learning with Temporal Dependence

Federico Giannini; Sandro D'Andrea; Emanuele Della Valle

cs.LGMar 9, 2026

Don't Look Back in Anger: MAGIC Net for Streaming Continual Learning with Temporal Dependence

Authors: Federico Giannini, Sandro D'Andrea, Emanuele Della Valle

Organizations: DEIB - Politecnico di Milano

Abstract

Concept drift, temporal dependence, and catastrophic forgetting represent major challenges when learning from data streams. While Streaming Machine Learning and Continual Learning (CL) address these issues separately, recent efforts in Streaming Continual Learning (SCL) aim to unify them. In this work, we introduce MAGIC Net, a novel SCL approach that integrates CL-inspired architectural strategies with recurrent neural networks to tame temporal dependence. MAGIC Net continuously learns, looks back at past knowledge by applying learnable masks over frozen weights, and expands its architecture when necessary. It performs all operations online, ensuring inference availability at all times. Experiments on synthetic and real-world streams show that it improves adaptation to new concepts, limits memory usage, and mitigates forgetting.

Figures & tables

Figure 1: Forward step on the feature vector

X_{t}

during the ensemble phase. For simplicity, only one Mask option is shown to represent both MaskFineTune and MaskRandom . They build weights (GRU M , OUT M ) by multiplying the frozen cGRU weights by masks in

(0,1)

. Masks are obtained by applying a sigmoid to learnable values. Expand expands the GRU layer with new learnable weights. It applies learnable masks to the frozen cGRU weights, and frozen masks equal to 1 to the new learnable weights. It obtains GRU E and OUT E . The opacity of the masked weights increases as the mask approaches 1.

		AirQuality		PowerConsumption		SRW		Weather
		start	end	start	end	start	end	start	end
100% prec 100% rec	ARF	0.07±0.02	0.08±0.02	0.25±0.05	0.29±0.05	0.24±0.02	0.33±0.01	0.26±0.05	0.29±0.04
	ARF T	0.21±0.05	0.26±0.04	0.22±0.06	0.28±0.06	0.73±0.00	0.73±0.00	0.41±0.06	0.44±0.06
	cGRU	0.33±0.05	0.39±0.05	0.59±0.07	0.70±0.04	0.69±0.05	0.81±0.03	0.50±0.06	0.61±0.06
	cPNN	0.35±0.03	0.35±0.03	0.66±0.02	0.71±0.02	0.61±0.05	0.68±0.04	0.52±0.05	0.56±0.05
	MAGIC	0.42±0.03	0.47±0.03	0.68±0.04	0.77±0.02	0.71±0.04	0.82±0.03	0.56±0.04	0.67±0.03

Table I : Results of the prequential evaluation with Cohen’s Kappa . For each data source and detector, the mean and standard deviation over 50 configurations are shown. Bold indicates best models. ARF T and cGRU excel on SRW. MAGIC Net leads on real data, with cPNN slightly worse at start on PowerConsumption and Weather. H1 is supported on real data.

Panel of Figure 2 — Figure 2: Average Cohen’s Kappa over time on 50 configurations of real data, using a concept drift detector with 100% precision and recall. Scores are first averaged per concept, then across configurations. MAGIC Net outperforms others across the stream. On PowerConsumption and Weather, cPNN starts slightly worse. MAGIC Net is clearly best at concept ends.

Panel of Figure 3 — Figure 3: Average sizes in MB on the 50 configurations of the real data sources over the detections with 70% precision and 100% recall. MAGIC Net reduces the size as new detections arise by deciding when to expand the architecture. H3 is proven.

		AirQuality		PowerConsumption		SRW		Weather
		AVG	BWT	AVG	BWT	AVG	BWT	AVG	BWT
100% prec 100% rec	ARF	0.01±0.01	-0.07±0.02	0.07±0.02	-0.35±0.08	0.09±0.02	-0.36±0.03	0.04±0.01	-0.18±0.04
	ARF T	0.14±0.04	-0.19±0.05	0.09±0.02	-0.31±0.12	0.73±0.00	-0.00±0.00	0.20±0.05	-0.24±0.06
	cGRU	0.06±0.02	-0.48±0.06	0.18±0.05	-0.86±0.08	0.20±0.05	-0.85±0.07	0.11±0.02	-0.68±0.06
	cPNN	0.31±0.05	-0.09±0.05	0.51±0.09	-0.32±0.15	0.59±0.10	-0.21±0.12	0.42±0.09	-0.19±0.09
	MAGIC	0.41±0.06	-0.11±0.05	0.55±0.10	-0.35±0.15	0.64±0.11	-0.26±0.14	0.49±0.07	-0.24±0.08

Table II : Results of the CL evaluation considering the Cohen’s Kappa . SML models and cGRU suffer from forgetting, while cPNN and MAGIC Net are robust. MAGIC Net is the best-performing model. H2 is almost always verified.

Continual Learning

Continual Learning studies how models acquire knowledge from changing data streams while retaining or selectively revising what they learned before.

See all 294 papers

A Practical Guide to Streaming Continual Learning
Mar 2, 2026Andrea Cossu, Federico Giannini, Giacomo Ziffer +5Continual Learning Streaming Algorithms
Computer Science Department, University of Pisa, Largo Bruno Pontecorvo, 3, Pisa, 56127, Italy · DEIB, Politecnico di Milano, Via Giuseppe Ponzio, 34, Milano, 20133, Italy · Motus ml, Milano, Italy +2
cPNN: Continuous Progressive Neural Networks for Evolving Streaming Time Series
Mar 3, 2026Federico Giannini, Giacomo Ziffer, Emanuele Della ValleOnline Continual Learning Concept Drift
DEIB, Politecnico di Milano, Milano, Italy
Streaming Continual Learning for Unified Adaptive Intelligence in Dynamic Environments
Mar 2, 2026Federico Giannini, Giacomo Ziffer, Andrea Cossu +1Continual Learning Online Continual Learning
DEIB, Politecnico di Milano, Milano, 20133, Italy · Computer Science Department, University of Pisa, Pisa, Italy

Mar 2, 2026J/K move · Enter open · S save

A Practical Guide to Streaming Continual Learning

Andrea Cossu, Federico Giannini, Giacomo Ziffer, Alessio Bernardo+4

Computer Science Department, University of Pisa, Largo Bruno Pontecorvo, 3, Pisa, 56127, Italy · DEIB, Politecnico di Milano, Via Giuseppe Ponzio, 34, Milano, 20133, Italy · Motus ml, Milano, Italy+2

Continual Learning Streaming Algorithms

Open paper AskPDF arXiv

Continual Learning (CL) and Streaming Machine Learning (SML) study the ability of agents to learn from a stream of non-stationary data. Despite sharing some similarities, they address different and complementary challenges. While SML focuses on rapid adaptation after changes (concept drifts), CL aims to retain past knowledge when learning new tasks. After a brief introduction to CL and SML, we discuss Streaming Continual Learning (SCL), an emerging paradigm providing a unifying solution to real-world problems, which may require both SML and CL abilities. We claim that SCL can i) connect the CL and SML communities, motivating their work towards the same goal, and ii) foster the design of hybrid approaches that can quickly adapt to new information (as in SML) without forgetting previous knowledge (as in CL). We conclude the paper with a motivating example and a set of experiments, highlighting the need for SCL by showing how CL and SML alone struggle in achieving rapid adaptation and knowledge retention.

Abstract

Continual Learning

Explore similar work

A Practical Guide to Streaming Continual Learning

cPNN: Continuous Progressive Neural Networks for Evolving Streaming Time Series

Streaming Continual Learning for Unified Adaptive Intelligence in Dynamic Environments