CLOPS: Quantum, Continual Learning & Avatar Control

• CLOPS is a context-dependent term with distinct definitions in quantum computing, continual learning for physiological signals, and vision-driven avatar motion generation.
• In quantum computing, CLOPS measures system-level execution speed by integrating classical orchestration with quantum circuit operations to complement metrics like quantum volume and qubit count.
• In continual learning and avatar control, CLOPS leverages replay strategies and egocentric vision to address catastrophic forgetting and enable natural motion behaviors, respectively.

to=arxiv_search.search 天天中彩票是 json {"query":"CLOPS arXiv", "max_results": 10, "sort_by":"relevance"}Japgollyassistant to=arxiv_search.search to=arxiv.search 北京赛车开 彩神争霸有ை 大发快三有json {"query":"CLOPS arXiv", "max_results": 10, "sort_by":"relevance"} to=search_arxiv 彩神争霸是 】【。】【”】【json {"query":"CLOPS", "max_results": 10, "sort_by":"relevance"} CLOPS is a context-dependent acronym used in several technically unrelated research literatures. In quantum computing, it denotes Circuit Layer Operations Per Second, a speed benchmark for near-term quantum computers that measures how quickly a system can execute layers of parameterized quantum volume circuits while accounting for both quantum hardware and classical control overhead (Wack et al., 2021). In continual learning for clinical time series, it denotes Continual Learning of Physiological Signals, a replay-based strategy for non-i.i.d. physiological data streams that couples replay buffers with learned task-instance parameters and uncertainty-driven rehearsal (Kiyasseh et al., 2020). In embodied character control, it denotes a vision-driven avatar motion system in which a human avatar uses only egocentric vision to perceive its surroundings and navigate (Diomataris et al., 23 Sep 2025). The acronym therefore has no single domain-independent definition.

1. Domain scope and disambiguation

The main uses of “CLOPS” represented in current arXiv literature are summarized below.

Usage	Domain	Core meaning
Circuit Layer Operations Per Second	Quantum computing	A speed benchmark for quantum systems
Continual Learning of Physiological Signals	Continual learning, clinical ML	A replay-based continual learning strategy
CLOPS avatar system	Avatar motion generation	A vision-driven avatar using egocentric perception

These usages are not methodological variants of one another. One concerns end-to-end hardware throughput, one concerns catastrophic forgetting under non-i.i.d. training streams, and one concerns perception-driven full-body control. This suggests that the term must be disambiguated from local context rather than from the acronym alone.

2. Circuit Layer Operations Per Second in quantum computing

In the quantum-computing literature, CLOPS measures how many layers of a “quantum volume” circuit a quantum computing system can execute per second, incorporating both quantum hardware speed and the interaction with classical control and processing systems (Wack et al., 2021). The benchmark is defined by

$$\text{CLOPS} = \frac{M \times K \times S \times D}{\textrm{time\_taken}}$$

where $M$ is the number of circuit templates, $K$ is the number of parameter updates per template, $S$ is the number of shots, and $D=\log_2(\mathrm{QV})$ is the number of quantum-volume layers. The prescribed protocol uses $M=100$, $K=10$, and $S=100$. Time starts before the first submission and stops when all results are returned, so circuit generation, parameter updates, runtime compilation, queueing, device initialization, execution, and data transfer are all included.

This formulation was introduced to complement the other two attributes identified for near-term quantum computers: quality, measured by quantum volume, and scale, measured by qubit count (Wack et al., 2021). The point of CLOPS is therefore not raw gate timing in isolation, but system-level responsiveness under parameterized, iterative workloads such as variational algorithms. The same source emphasizes that improvements in runtime compilation, orchestration, and classical control can increase CLOPS even without changes to the quantum processor itself.

Representative values reported for IBM systems under the prescribed protocol were 1419 for ibmq_bogota, 951 for ibmq_toronto, and 753 for ibmq_brooklyn (Wack et al., 2021). A detailed time breakdown for ibmq_brooklyn illustrates the benchmark’s systems perspective: total time was 663.6 s, of which circuit execution accounted for 2.0 s, circuit delays for 25.0 s, and run-time compilation plus data transfer for 636.6 s. A common misconception is therefore that CLOPS is merely a proxy for gate speed; the benchmark was explicitly designed to capture the combined classical–quantum execution path.

3. CLOPS as a systems and scheduling constraint

Subsequent quantum-systems work treats CLOPS as a hardware attribute and studies how workload decomposition and microarchitecture affect realized throughput. One study on IBM’s Quantum Cloud states that “CLOPS is a metric which captures the execution speed of a quantum hardware” and further clarifies that CLOPS is a hardware attribute — not an application workload metric but a device-side capability indicator (Khare et al., 2023). Its Table I reports CLOPS values of 2.3k for ibm hanoi, 2.6k for ibm oslo, 2.6k for ibm nairobi, 2.9k for ibmq perth, 2.8k for ibmq manila, 2.5k for ibmq belem, and 2.7k for ibmq lima.

That study evaluates two splitting techniques for hybrid quantum-classical workloads: circuit parallelization, which splits one large circuit into multiple smaller ones, and data parallelization, which splits a large number of circuits into smaller batches run on different hardware (Khare et al., 2023). The reported trade-offs are explicitly multi-objective. For VQE, circuit cutting is ~39% better in ground state estimation than the uncut version, but increases resource foot-print and classical post-processing overhead. For QSVM, data parallelization with reduced feature set yields upto 3x improvement in quantum workload execution time and reduces quantum resource use by 3x, while providing comparable accuracy. Error mitigation can improve the accuracy by ~7% and resource foot-print by ~4% compared to the best case among the considered scenarios. The same analysis uses the qubit-time product,

$\text{Qubit-Time Product} = Q \times t,$

to make explicit that reduced wall-clock time does not necessarily imply reduced total resource consumption.

A more architectural development is HiMA, a hierarchical quantum microarchitecture for a 72-qubit tunable superconducting quantum processing unit serving a public quantum cloud computing platform (Zhou et al., 2024). HiMA is based on three core elements: discrete qubit-level drive and readout, a process-based hierarchical trigger mechanism, and multiprocessing with a staggered triggering technique to enable efficient quantum process-level parallelism. The design is reported as capable of expanding to 6144 qubits through three-layer cascading. In benchmarking, HiMA achieves up to a 4.89x speedup under a 5-process parallel configuration and reaches up to 43,680 CLOPS; the same source states that this is the highest CLOPS across all publicly available platforms (Zhou et al., 2024). Reported single-process CLOPS is 12,304, while five-process total CLOPS is 43,680, with 94.48% speedup efficiency and QPU load average rising from 16% to 66%.

Taken together, these results position quantum-computing CLOPS as a benchmark that is simultaneously architectural, orchestration-sensitive, and workload-sensitive. High CLOPS depends not only on the quantum processor, but also on trigger design, classical scheduling, data movement, and queue behavior.

4. CLOPS as Continual Learning of Physiological Signals

In machine learning for clinical time series, CLOPS denotes Continual Learning of Physiological Signals, a replay-based continual learning strategy designed for physiological signal data streams that are non-i.i.d. because of temporal dependence, varying sensor modalities, and heterogeneous patient or environment populations (Kiyasseh et al., 2020). The method was introduced to address catastrophic forgetting in clinical settings, where data are streamed temporally and from a multitude of physiological sensors.

The two core mechanisms are importance-guided buffer storage and uncertainty-based buffer acquisition (Kiyasseh et al., 2020). For each instance $x_{i\mathcal T}$ in task $M$0, CLOPS maintains a scalar task-instance parameter $M$1. These parameters are updated by gradient descent and are used both for loss weighting and for deciding which samples enter the replay buffer. For the current task $M$2, the weighted objective is

$M$3

Because

$M$4

instances with higher loss get their $M$5 pushed lower, so lower $M$6 acts as a difficulty proxy. Storage is not based on the terminal value alone; instead, instances are ranked by the integrated score

$M$7

For replay acquisition, CLOPS uses MC-Dropout and the BALD criterion,

$M$8

The method was evaluated on three large ECG datasets under three continual learning scenarios: Class-IL, Time-IL, and Domain-IL (Kiyasseh et al., 2020). The reported CLOPS scores were as follows.

Scenario	Average AUC	BWT
Class-IL	0.796	0.053
Time-IL	0.834	-0.018
Domain-IL	0.731	-0.011

The paper states that in three continual learning scenarios based on three publically-available datasets, CLOPS can outperform the state-of-the-art methods, GEM and MIR (Kiyasseh et al., 2020). The detailed comparisons reported in the summary show that in Class-IL, CLOPS exceeds MIR and GEM; in Domain-IL, it again exceeds both; and in Time-IL, its average AUC is close to MIR while being described as showing strong forward transfer. The same study also reports robustness to task order.

A distinctive feature of this usage is interpretability. Task-instance parameters are used to quantify task difficulty and similarity, with task similarity derived from Gaussian fits to the $M$9-distributions and a Hellinger-distance-based similarity measure (Kiyasseh et al., 2020). The source argues that this quantification yields insights into both network interpretability and clinical applications, where task difficulty is poorly quantified. It further reports that curricula based on these scores can increase constructive interference.

5. CLOPS in vision-driven avatar motion generation

In avatar motion generation, CLOPS denotes a system introduced as “the first human avatar that solely uses egocentric vision to perceive its surroundings and navigate” (Diomataris et al., 23 Sep 2025). The motivating claim is that the generation of human-like avatar behavior requires human-like perception, and the method therefore uses egocentric vision rather than privileged task-specific sensing.

The system is explicitly decoupled into two learned components (Diomataris et al., 23 Sep 2025). The first is a motion prior $K$0, a conditional Variational Autoencoder trained on a large motion capture dataset. It autoregressively synthesizes natural full-body motion conditioned on a target pose of the head and outputs a motion chunk $K$1. The second is a visual policy $K$2, trained using Q-learning to map egocentric visual input to high-level control commands for the motion prior. The overall decomposition is

$K$3

The egocentric input is specified as $K$4, consisting of a depth map, semantic segmentation, and a goal mask, rendered from the avatar’s head with a 130° FOV (Diomataris et al., 23 Sep 2025). The policy chooses among a discrete set of head-pose deltas, and training uses Double Q-learning and prioritized experience replay. The reward is defined as

$K$5

The reported behaviors are qualitative and quantitative. The avatars walk such that they avoid obstacles present in their visual field, frequently orient their head and body toward the direction of travel, look around to find the goal, and keep goals in view while moving (Diomataris et al., 23 Sep 2025). A sensor-placement study reported that when the avatar’s eye is placed facing backward, the avatar tends to walk backwards to match the new perception-action coupling. Quantitatively, policies trained in one environment transfer to unseen environments with success rates above 66% across all held-out test scenes. In the cited comparison on test scene S5, TriloAvatar achieves a 68% success rate versus 48% for EgoGen, with lower foot-skating and better collision behavior; in an oracle case where the goal location is also provided, the system achieves 100% success (Diomataris et al., 23 Sep 2025).

This usage of CLOPS is unrelated to both the quantum benchmark and the continual-learning method. Here the acronym identifies a decoupled perception-to-motion architecture for embodied control.

6. Conceptual boundaries and recurring misconceptions

The three principal usages of CLOPS differ not only in application area but also in the level of abstraction at which the term operates. In quantum computing, CLOPS is a benchmark: it measures end-to-end execution speed for parameterized quantum-volume workloads and explicitly includes classical orchestration, compilation, and data-transfer overhead (Wack et al., 2021). In clinical continual learning, CLOPS is an algorithmic strategy: it combines replay, learned task-instance parameters, and uncertainty-based acquisition to mitigate destructive interference under non-i.i.d. physiological streams (Kiyasseh et al., 2020). In avatar motion generation, CLOPS is an embodied control system: it maps egocentric vision to high-level control commands consumed by a motion prior (Diomataris et al., 23 Sep 2025).

Several misunderstandings recur across these literatures. First, quantum-computing CLOPS is not simply a measure of gate speed; the original formulation and later analyses emphasize the role of classical components and cloud-level scheduling (Wack et al., 2021, Khare et al., 2023). Second, continual-learning CLOPS is not a generic replay baseline; its defining feature is the end-to-end trainable task-instance parameterization used for both loss weighting and buffer storage (Kiyasseh et al., 2020). Third, avatar CLOPS does not rely on omniscient sensing: the system is introduced precisely as one that uses only egocentric vision, with human-like behaviors emerging without explicit behavioral shaping (Diomataris et al., 23 Sep 2025).

This suggests that references to “CLOPS” are not self-disambiguating. In practice, the surrounding vocabulary identifies the intended sense: quantum volume, qubits, backends, and cloud throughput indicate the quantum metric; ECG, AUC, BWT, replay buffer, and task-instance parameters indicate continual learning of physiological signals; and egocentric vision, head target pose, motion prior, and Q-learning indicate the avatar-control system.