Analog in-memory chips compute neural-network layers inside the memory arrays that store the weights. This saves a great deal of energy, but every weight is slightly wrong: it fluctuates on each read, drifts over time, and differs from chip to chip. The field judges these errors by classification accuracy on fixed test sets. It also reports that moderate noise can help, acting like built-in dropout. Much less is known about what happens when the network is a controller running in a feedback loop that serves a person. One example is the model inside a video call that decides, every 60 milliseconds, how much bandwidth to use. For that model the right question is how much worse the call looks and sounds to the person on it, not how far the output drifts from the digital answer.
We propose to measure analog weight errors in units of predicted mean opinion score (MOS, a 1–5 rating of perceived audio or video quality). The testbed is learned bandwidth estimators for real-time calls. Our central hypothesis is simple. Sending too much data causes packet loss and frozen frames, which people hate. Sending a little too little costs only a slight, barely noticeable drop in sharpness. If so, perceived quality should tolerate zero-mean, per-step read noise far beyond the "elbow" seen in classification. It should be highly sensitive to persistent errors that push the estimate upward. The practical consequence is that analog compensation for such controllers should be deliberately lopsided, trading a small, controlled underestimate for protection against overshoot.
Core Claim
Write the deployed bandwidth estimate as
$$\hat{a}_t = a_t\,(1 + \beta + \epsilon_t),$$
where $a_t$ is the estimate the digital policy would produce at step $t$ and $\beta$ is a persistent relative bias from drift, update asymmetry or device variation. The term $\epsilon_t$ is per-step, zero-mean read noise with spread $\sigma$.
We claim three things. First, predicted MOS loss grows much faster in $\beta$ for $\beta>0$ than for $\beta<0$. Second, MOS loss is nearly flat in $\sigma$ up to a threshold. The flatness is expected because the jitter buffer and the congestion feedback average over many 60 ms steps, much as stochastic pulse streams average out update noise in Software-Level Accuracy Using Stochastic Computing With Charge-Trap-Flash Based Weight Matrix. Third, the tolerable noise level for real device models is therefore set by bias, not by variance. A compensation rule that estimates $\beta$ and shifts the output slightly toward underestimation should recover most of the digital policy's MOS.
The first two claims rest on two sources. The perceptual diminishing-returns results give the saturation side, in QoE-Based Resource Allocation for Multi-Cell NOMA Networks and Managing power in heterogeneous multicore systems. The observation in ACM MMSys 2024 Bandwidth Estimation in Real Time Communications Challenge that overshooting capacity degrades video quality gives the steep side.
What is new is not "run a policy on analog hardware." It is the claim that perceptual, closed-loop scoring changes which hardware errors matter, and that compensation should follow that asymmetry. Prior analog work compensates drift symmetrically and scores by accuracy. The main example is the global drift compensation in Toward Software-Equivalent Accuracy on Transformer-Based Deep Neural Networks With Analog Memory Devices, which scores GLUE.
Research Program
We train bandwidth-estimation policies offline on the MMSys 2024 challenge data, using its MOS-predictor-based reward, following the challenge setup ACM MMSys 2024 Bandwidth Estimation in Real Time Communications Challenge. That data comprises 18,859 real Microsoft Teams calls across six behavior policies and 9,405 emulated calls with ground-truth capacity. The policy consumes the challenge's 150-dimensional observation of receive rates, jitter and loss over 60 ms and 600 ms windows. It is small, a few dense or recurrent layers, which keeps the study fast and makes every hardware effect traceable.
We then map the trained weights onto simulated analog arrays with the IBM Analog Hardware Acceleration Kit (aihwkit). Two device models are used. The first is phase-change memory with programming noise, read noise and power-law drift, as used by Toward Software-Equivalent Accuracy on Transformer-Based Deep Neural Networks With Analog Memory Devices. The second is HfOx resistive memory, as used by Assessing the Performance of Analog Training for Transfer Learning.
Closed-loop evaluation replays the ground-truth capacity traces from the emulated calls through a trace-driven network emulator. The policy's decisions change queueing, loss and delay, and the challenge MOS predictors score the resulting call. If the challenge's own emulation platform cannot be reused outside the challenge, we will build this replay harness ourselves. That is the main engineering cost.
The analysis has three parts. First, we separate the hardware errors. Using aihwkit's configurable noise, we sweep read noise alone, drift alone at several simulated ages, and fixed device-to-device variation alone. For each sweep we record MOS, packet loss, dropped frames, and the overshoot ratio, meaning the fraction of steps where the estimate exceeds true capacity. Second, we fit the empirical MOS loss as a function of the induced $\beta$ and $\sigma$ to test the asymmetry claim directly. Third, we build asymmetric bias compensation. A few reference weights (or the network's output on a fixed calibration input) are read periodically to estimate $\beta$. The output is then scaled by $1/(1+\hat\beta)$ times a small margin $(1-\delta)$, where $\delta$ is chosen from the fitted asymmetric loss curve. We also test hardware-aware training, meaning noise injected into the forward pass during offline training, on its own and combined with the compensation.
A secondary question concerns regularization. Earlier work finds that analog noise helps generalization, in Towards Exact Gradient-based Training on Analog In-memory Computing and Surrogate gradients for analog neuromorphic computing. We ask whether training with analog-realistic noise narrows the gap between emulated and real-call behavior. That gap is the one that hurt R3Net on 3G links in Reinforcement learning for bandwidth estimation and congestion control in real-time communications. We treat this as secondary and report it honestly, because a digital-noise control may explain the entire effect.
On-chip online adaptation with chopped Tiki-Taka is deliberately left out of scope; it is a follow-on only if the bias story holds.
Baselines
The naive reference is the digital FP32 policy trained without noise and deployed on the simulated analog array with no compensation. It shows how bad uncorrected hardware is in perceptual terms.
The upper reference is the same FP32 policy run digitally in the same emulation harness. For context, the winning challenge model Schaferct scored 8.93/10 on the global testbed in ACM MMSys 2024 Bandwidth Estimation in Real Time Communications Challenge. We cannot rerun that testbed, so all our comparisons are relative within our harness.
The classical baseline is an Unscented Kalman Filter bandwidth estimator of the kind used in production by Reinforcement learning for bandwidth estimation and congestion control in real-time communications. On 3G it reported 2.22% packet loss and 6.5% dropped frames, against 3.11% and 11.2% for R3Net. It needs no neural hardware at all, so any analog policy that loses to it on MOS has no reason to exist.
The strongest existing analog method has two parts: hardware-aware training with realistic noise injection, plus global symmetric digital drift compensation. Both follow Toward Software-Equivalent Accuracy on Transformer-Based Deep Neural Networks With Analog Memory Devices, which reports aggregate GLUE within 1.29% of software after one simulated month of drift. No reported result was found on Lacuna for this method on bandwidth estimation or any MOS metric. We give it the same training data, the same number of noise-injection training runs, and the same calibration reads as our method.
The key ablation replaces our asymmetric compensation with symmetric compensation that uses the identical $\hat\beta$ estimate. It also compares a fixed conservative offset with no drift tracking. Together these isolate whether the asymmetry, the tracking, or both do the work.
Our method beats the strongest baseline if two things hold. It must retain at least the same fraction of digital MOS at one simulated month of drift. It must also show a statistically lower overshoot ratio and dropped-frame rate across the held-out emulated calls, with paired comparisons on identical traces.
Why not just…?
Why not just inject digital Gaussian noise or dropout instead of relying on hardware noise? For the regularization question, this is a real threat, and the file offers no evidence that analog noise does anything matched digital noise cannot. We therefore run a digital control whose injected noise matches the aihwkit read-noise spread. If it closes the gap between emulated and real calls as well as analog-realistic noise does, we will say regularization comes from noise statistics, not hardware. The main contribution does not depend on this outcome, because it concerns how to deploy on analog hardware, where the noise is present whether one wants it or not.
Why not just keep the Unscented Kalman Filter? It is cheap and was more robust on 3G than R3Net in Reinforcement learning for bandwidth estimation and congestion control in real-time communications. However, it targets capacity rather than perceived quality. The MMSys challenge found that capacity-focused estimation does not reliably raise video MOS ACM MMSys 2024 Bandwidth Estimation in Real Time Communications Challenge. It is included as a baseline in the same harness, and the results will show directly whether a MOS-trained policy keeps its advantage once hardware errors are added.
Why not just quantize to INT8 on a digital accelerator? For a policy this small, deterministic low-precision digital inference is a strong, possibly sufficient, option. The energy argument for analog hardware on this one controller is weak. We include an INT8 policy as a reference point for MOS. We frame the contribution as a method for scoring and compensating analog errors in closed perceptual loops, with bandwidth estimation as a well-instrumented test case, not as a claim that this controller needs analog hardware. The file gives no energy figures for analog execution of this workload, and we do not claim any.
Why not just bias the output downward by a constant? This is the simplest form of our idea, and it may be enough if drift is slow relative to recalibration. The fixed-offset ablation tests it directly. Drift tracking earns its place only if its advantage grows with simulated device age.
Decisive Checks
The first check is the asymmetry itself. Across the held-out emulated calls, we inject a controlled output bias of equal magnitude in each direction. If MOS loss at $\beta=+0.1$ is not clearly larger than at $\beta=-0.1$, the central premise fails and the compensation design is moot.
The second check compares units. Classification work places the stability elbow near 15% weight noise Assessing the Performance of Analog Training for Transfer Learning. We locate the read-noise level at which MOS begins to fall and the level at which the output's mean-squared error doubles. If the MOS threshold is not substantially higher than the error-based threshold, perceptual scoring adds nothing over existing metrics.
The third check tests compensation after drift. At simulated ages from one hour to one month, asymmetric compensation must beat symmetric compensation with the same $\hat\beta$ on the overshoot ratio and dropped frames. It must do so without losing more than a small, pre-registered amount of mean MOS.
The fourth check is the matched digital-noise control on the gap between emulated and real behavior, which settles the regularization question either way. The real-call half of this check needs a trace-driven proxy. We build it by replaying received-rate and loss patterns from the 18,859-call real dataset as capacity traces and scoring the policy on them.
References
- Software-Level Accuracy Using Stochastic Computing With Charge-Trap-Flash Based Weight Matrix
- QoE-Based Resource Allocation for Multi-Cell NOMA Networks
- Managing power in heterogeneous multicore systems
- ACM MMSys 2024 Bandwidth Estimation in Real Time Communications Challenge
- Toward Software-Equivalent Accuracy on Transformer-Based Deep Neural Networks With Analog Memory Devices
- Assessing the Performance of Analog Training for Transfer Learning
- Towards Exact Gradient-based Training on Analog In-memory Computing
- Surrogate gradients for analog neuromorphic computing
- Reinforcement learning for bandwidth estimation and congestion control in real-time communications