Carousel memory

Carousel memory refers to two distinct concepts: a historical type of computer secondary storage from the late 1950s, and a modern hierarchical memory management strategy for continual learning in machine learning.

1. Historical Computer Hardware

Carousel memory is a type of secondary storage for computers, created by Swedish computer engineers Erik Stemme and Gunnar Stenudd. It was first shown at an exhibition in Paris in 1958. The device was manufactured by the Swedish company Facit AB under the model designation FACIT ECM 64.

Design and Operation

Rather than using a single long magnetic tape, the carousel memory employed 64 small tape rolls, each containing 9 meters of 1.6 cm (5/8‑inch) wide magnetic tape. Each roll was divided into eight channels. The tape moved at a speed of 5 m/s. To access a particular roll, the carousel rotated so that the desired roll was positioned at the bottom; a counterweight at the free end of the tape then guided it into a mechanism containing a read/write head. After reading or writing, the tape was rewound.

Specifications

• Storage capacity: 2,560 kilobytes
• Average seek time: 2 seconds
• Peak transfer rate: 182,044 bits/s (using eight parallel lines at 22,756 bits/s per line)
• Tape medium: 5/8‑inch wide, 0.05 mm thick Mylar (3M Co. type 188)
• Storage density: 8 bits/mm
• Control system: Transistor-based
• Power requirement: Three‑phase 380 V, 50 Hz; 300 W standby, 750 W active
• Signaling: Eight parallel –20 V to 0 V, 5 μs pulses

Both the carousel assembly and the individual spools were replaceable.

Historical Significance

The first delivery of a Facit EDB 3 computer (to ASEA in Västerås, Sweden, in 1958) used the carousel memory FACIT ECM 64. The device represents an early electromechanical approach to random‑access secondary storage, predating the dominance of magnetic disk drives.

2. Machine Learning Concept (CarM)

In the field of continual learning (CL), Carousel Memory (CarM) is a hierarchical episodic memory management strategy proposed in the 2021 paper "Carousel Memory: Rethinking the Design of Episodic Memory for Continual Learning" (arXiv:2110.07276).

Motivation

Continual learning aims to train neural networks on a continuous stream of tasks without catastrophic forgetting. Traditional episodic memory (EM) methods store a subset of past samples in a small, fast on‑device RAM buffer. However, the limited size of this buffer constrains accuracy. CarM exploits the observation that mobile and IoT devices possess a deep memory hierarchy: small, fast RAM (50–150 ns access) and much larger, slower internal storage (25–250 μs access, often orders of magnitude larger).

Design

CarM stores as many observed samples as possible on the device's internal storage (rather than discarding samples that overflow from EM) and dynamically swaps samples between the fast EM and the slow storage during training. The key components are:

• Swap worker: An asynchronous process that issues I/O requests to bring samples from storage into memory, overlapping I/O latency with computation.
• Gate function: A scoring mechanism that decides which in‑memory samples to replace with in‑storage samples. Three policies are proposed:
  1. Random – randomly selects samples for replacement.
  2. Entropy – prefers to replace correctly predicted samples with low entropy (high confidence) and retain incorrectly predicted samples with low entropy.
  3. Dynamic – uses Random for the first half of training passes and Entropy for the second half.
• Asynchronous sample retrieval: Hides the latency of storage access by running retrieval in parallel with subsequent training steps.

Performance

CarM is complementary to existing EM‑based CL methods. Evaluated with seven popular CL methods (ER, iCaRL, TinyER, BiC, GDumb, DER++, RM) on CIFAR and ImageNet datasets, CarM improved final average accuracy by up to 28.4 % while retaining the same training efficiency. It also allowed methods to maintain accuracy with much smaller EM sizes.

References

1. Pär Rittsel (2002), Valsen inledde skivans historia, International Data Group.
2. FACIT electronics AB (1961), TPU 610301 FACIT ECM 64 Karusellminne, Preliminär teknisk beskrivning.
3. FACIT electronics AB (1963), Facit Electronic presenterar FACIT ECM 64 Magnetbandsminne av karuselltyp med direkt access.
4. Antiquus S. Hippocampus et al. (2021), Carousel Memory: Rethinking the Design of Episodic Memory for Continual Learning, arXiv:2110.07276.
5. Karlqvist, Olle (March 1962), Applications to the magnetic tape storage unit, facit ECM64 (the Carousel Memory), BIT Numerical Mathematics 2(1): 16–20.