The Chip That Trains on Its Own Imperfections

A chip is running hot. Not in temperature, but in the thermodynamic sense. Every operation dissipates energy into noise and variability, stored weights drift from their intended values, and each pulse lands somewhere close to but not exactly on target. This is not a defective chip. This is simply what analog memory hardware does. In a digital world, 1 + 1 is always 2; in the analog world of AIMC, 1 + 1 might be 2.04 today and 1.98 tomorrow. The hardware is, in a practical sense, almost alive, and its behavior shifts with temperature, wear, and the statistical quirks of individual memory cells. And for years, that basic fact has kept one of the most consequential frontiers in computing locked behind a wall of compounding error.

The frontier in question is on-chip training. It means a neuromorphic computing system learns directly on the analog hardware where it will eventually run, rather than on digital supercomputers before transfer. The prize is enormous: training systems that operate at a fraction of the energy cost of digital hardware, running on chips embedded at the edge rather than in distant data centers. Achieving that goal has been stubbornly difficult, not because researchers lack ideas, but because the physics of analog devices keep getting in the way.

A group of researchers at Seoul National University has now published a method they call SALMON, short for Self-Adaptive Learning Model on Neuromorphic Hardware. It turns this problem inside out. Rather than trying to suppress analog hardware flaws from the outside, their approach teaches the network to adapt around those flaws from within. The results appear in IEEE Transactions on Neural Networks and Learning Systems. They are notable both for their accuracy and for what they reveal about the deeper relationship between hardware imperfection and learning.


The Problem That Would Not Go Away

To understand why on-chip training is so difficult, it helps to understand what makes analog in-memory computing attractive in the first place.

Why Analog Looked So Compelling

Conventional digital chips perform matrix-vector multiplications, the mathematical backbone of deep learning, by shuttling data back and forth between memory and processing cores. That shuttling consumes a large share of the total energy budget. Analog in-memory computing, or AIMC, collapses this by storing neural network weights directly in resistive memory arrays and performing the multiplication in place, using Ohm’s law and Kirchhoff’s current law. The weights live in the conductance of the memory cells. Rather than sequential multiply-accumulate steps, the hardware executes the operation in a single physical step, effectively constant-time with respect to vector dimension under full parallelism. Digital architectures cannot easily replicate that speed and energy advantage.

Three Ways Analog Hardware Fights Back

The same physical properties that make AIMC fast, however, make it imprecise. Three categories of imperfection cause the most trouble.

The first is limited conductance resolution. An analog memory cell can only hold a finite number of conductance states, which is equivalent to having low numerical precision. Fewer states mean coarser weight updates and degraded learning.

The second is update asymmetry, which the paper captures through a parameter called the gamma level. When a pulse increases a weight versus decreases it, the resulting conductance change differs. The weight drifts differently depending on direction, which sends gradient-based optimization off course in ways that standard stochastic gradient descent never handled.

The third is device variation, both cycle-to-cycle variation within a single cell and device-to-device variation across the array. No two cells behave exactly alike. The same cell does not behave exactly the same way twice. Algorithms that assume spatially uniform hardware run into trouble when it is not.

Why Prior Solutions Fall Short

These three nonidealities interact and compound as a network grows deeper. Each layer introduces its own errors into both the forward pass and the backward pass. By the time gradients reach the early layers, they carry noise from every subsequent layer. Researchers have proposed workarounds: the Tiki-Taka algorithm uses two analog arrays per layer to handle asymmetry, and mixed-precision methods store weight updates in high-precision digital memory. Each approach carries its own costs in area, speed, or energy.

What these methods largely avoid is letting the network directly learn the idiosyncrasies of a specific chip instance in real time. None of them adapt to what that particular piece of silicon actually does, only to a statistical model of what the hardware might do.


Bridging the Gap Between Simulation and Silicon

Before examining how SALMON works, it is worth naming a structural problem that the field has long struggled to articulate cleanly: the sim-to-real gap.

The Gap That Breaks Deployment

When engineers develop new training algorithms for AIMC, they typically validate those algorithms in software simulation on a GPU or CPU. The simulation can be made arbitrarily precise. The physical chip cannot. Algorithms that perform well in simulation often degrade on actual hardware, because no simulation fully captures the noise signature of a specific piece of silicon. Every chip is slightly different, and every chip ages differently. A simulation might assume symmetric weight updates with Gaussian noise, while a real ReRAM device exhibits strongly skewed update steps that depend on its current conductance state. That discrepancy between model and reality is exactly where deployment failures tend to originate.

Live Adaptation vs. Modeled Noise

Hardware-aware training, a well-established approach in the field, attempts to close this gap by injecting simulated device noise into the training process while still running on a GPU. The training process learns robustness to that modeled noise. However, hardware-aware training guesses what the noise will look like based on a model. SALMON measures the actual noise of that specific piece of hardware, in real time, and adapts to it directly. Think of the difference between training a surgeon on a simulation and training them in the operating room. The latter produces adaptation that is specific, not generic.

This live-adaptive quality is part of what makes SALMON architecturally distinct from prior approaches, and it matters most in the devices where simulation-to-hardware gaps are largest.


The Self-Adaptive Network Architecture

The core contribution of SALMON is an architecture called SAnet, which stands for Self-Adaptive Network. SAnet pairs an analog backbone, built from ResNet-style residual blocks running on resistive processing units (RPUs), with digital attention blocks running on conventional digital processing units (DPUs). The two sides serve different purposes during training, and only one of them runs during inference.

Three-Tiered Learning: Muscle, Eyes, and Coordinating Signal

It helps to think of SAnet in three functional tiers, loosely analogous to roles in a coordinated system.

The first tier is the analog backbone, the muscle. It handles approximately 94.9% of total floating-point operations at low energy cost. However, it is imprecise. The noise and drift described above accumulate across every forward and backward pass, and every weight update lands near, but not exactly at, the intended conductance value.

The second tier is the digital attention blocks, the eyes. These blocks compute precise, clean gradients using full digital arithmetic. They see clearly what the analog backbone cannot, because they are not subject to the same physical noise. The cost is significant: digital units draw roughly 100 times more power per operation than analog units. Although the digital blocks handle a small fraction of total computation, they account for roughly 85% of total power consumption during training.

The third tier is the composite gradient, the coordinating signal. This is where analog and digital information merge into a single update direction that is more accurate than either source alone. The composite gradient is the mechanism through which SAnet adapts, and it transforms an ordinary hybrid architecture into a self-adaptive one.

The Trust Parameter: How Alpha Mediates the Synthesis

During training, both the analog backbone and the digital attention blocks are active. The composite gradient combines the noisy analog backbone gradient with a scaled version of the clean digital attention gradient. The scaling factor, called alpha in the paper, functions as a trust parameter. In practical terms, alpha answers a simple question: how much should the system trust physics versus arithmetic?

At high alpha, the system places substantial trust in the digital attention blocks. The composite gradient rotates toward the clean, precise direction those blocks compute, which is closer to the true loss minimum but further from the raw analog signal. At low alpha, the system forces the analog backbone to rely more heavily on its own noisy gradient, which preserves fidelity to the hardware’s native update dynamics but sacrifices directional accuracy.

The researchers found that performance peaks near alpha equal to one across all three device models tested. The best results come from roughly equal weighting of both sources. This is not simply averaging away noise. The gradient mathematics, discussed further below, confirm that the digital attention signal systematically biases the analog updates toward better directions rather than merely damping random fluctuations.

Over many training epochs, this repeated correction shapes the network’s intermediate feature representations. They grow less sensitive to the hardware’s imperfections, not because the hardware improved, but because the network learned to work within the hardware’s specific constraints.

During inference, the researchers remove the digital attention blocks entirely. Only the analog backbone runs, with no external digital processing units required. The network trained with digital assistance but does not depend on it once training is complete.


What the Gradient Mathematics Actually Shows

The researchers conducted a formal local gradient analysis to verify that the self-adaptation mechanism does what it appears to do, rather than simply averaging away noise through brute force.

Measuring the Beta Signal

They defined a quantity called beta, which measures the geometric relationship between the digital attention gradient and the analog backbone gradient at each of the three attachment points along the network. A positive beta indicates that the digital gradient reinforces the descent direction of the analog gradient. Beta is positive at all three attachment points. It also increases in magnitude as layer depth decreases, meaning the effect is strongest in the earlier layers where analog noise has had more passes to accumulate.

Quantifying the Alignment Gain

The researchers also defined an alignment gain, which measures how much closer the composite gradient points toward the ideal full-precision gradient compared to the analog gradient alone. Across all training epochs and all three attachment points, this alignment gain stayed consistently positive at the operating value of alpha. The digital attention blocks reliably steer the analog updates toward better directions.

There is a tradeoff, which the researchers quantified carefully. As alpha increases, the composite gradient rotates further from the pure analog backbone gradient, reducing a coherence measure between the two. But increasing alpha also increases amplification in the direction of the ideal full-precision reference. Test accuracy peaks near alpha equal to one across all three device models, suggesting the researchers operated near the optimal balance point.


Three Device Models and the Real Hardware They Represent

The researchers evaluated SALMON against three distinct analog device models representing different levels of physical realism, rather than testing against a single idealized model.

From Ideal to Nonideal

Device Model 1 is an ideal analog device using 32-bit floating-point precision for weight updates. Practical nonidealities from peripheral circuits, including ADC and DAC quantization effects, still affect forward and backward passes at this level. This model establishes a ceiling for what analog hardware can do in the best case.

Device Model 2 uses 10,000 conductance states and a fully parallel stochastic update scheme. The hardware maintains perfect update symmetry, but cycle-to-cycle and device-to-device conductance variations are active. This model represents the kind of analog hardware that careful engineering might achieve in the near term.

Device Model 3 is the most physically realistic. It uses empirical operational characteristics from actual devices, and this is where the industry relevance becomes most direct.

ECRAM and ReRAM: Two Very Different Bets

The two device technologies grounding Device Model 3 occupy different positions in the commercial landscape. Understanding that difference clarifies why SALMON’s results matter to engineers choosing between them.

Electrochemical RAM, or ECRAM, ranks among the most symmetric analog memory technologies currently in research. Its update behavior is relatively clean: the gamma level is low, conductance states are numerous, and the device does not fight the training algorithm as aggressively. The tradeoff is manufacturability. ECRAM remains largely confined to research fabs and has not yet achieved the scale or cost structure of mature semiconductor processes.

Resistive RAM, or ReRAM, is the opposite story. Foundries such as TSMC and Infineon already offer it commercially, integrated into products shipping today. However, ReRAM is noisier, more asymmetric, and more prone to device variation than ECRAM. For on-chip training purposes, its severe nonidealities have historically made it a difficult device to work with.

SALMON’s results on ReRAM are consequently among the most practically meaningful in the paper. Combined mixed-precision training and self-adaptation on ReRAM reaches 82.46% accuracy versus 73.06% without self-adaptation, a gain of more than 9 percentage points. In practice, this positions SALMON less as a laboratory curiosity and more as a bridge toward deployable on-chip training using already-available fabrication processes. For a device engineers can actually purchase today, that improvement is not a theoretical footnote.


The Schooling Phase: Paying an Early Energy Tax

The paper’s most practically useful contribution may be its strategies for reducing how often the digital attention blocks run during training.

The Cost of Digital Oversight

The blocks are the expensive part of SALMON. The analog backbone accounts for 94.9% of total floating-point operations, but the digital components account for roughly 85% of total power consumption. That asymmetry exists because digital units draw roughly 100 times more power per operation than analog units.

Two Strategies for Cutting That Cost

The sparse strategy activates the digital attention blocks only once every Ts training epochs. At Ts equal to 4, digital blocks run only one quarter as often. Accuracy drops by less than 5% compared to full activation across all device models tested, while digital component usage drops by 75%.

The warm-up strategy deserves closer attention for its conceptual framing. It activates the digital attention blocks intensively during the first Tw epochs, then turns them off for the remainder of training. This reframes digital assistance not as a permanent crutch but as a temporary instructor. The system pays an energy tax early on. The digital blocks educate the analog backbone by repeatedly injecting clean gradient signals during the period when the network’s representations are most malleable. Once that early shaping ends, the analog backbone carries the correction forward on its own. The system graduates to a pure-analog mode that costs almost nothing by comparison.

At Tw equal to 30 epochs out of 300 total, the accuracy degradation is at most 3.16%, while digital component usage drops by 90%. The analog backbone learned what it needed in the first 10% of training and carried that knowledge the rest of the way.

What the Numbers Yield

The combined mixed strategy uses warm-up followed by sparse activation, achieving further reductions. The sparse strategy at Ts equal to 4 cuts relative power consumption by a factor of approximately 40 compared to a fully digital training baseline. The warm-up strategy at Tw equal to 30 brings that factor to approximately 56. During inference, with digital blocks removed, the reduction reaches approximately 73 times the fully digital baseline.

In terms of computational throughput per watt, SALMON training achieves 4.2 times the TOPS/W of the fully digital baseline. That figure rises to 8.5 times for the sparse strategy and 11.4 times for the warm-up strategy. The ALPINE framework, a hardware-validated model for AIMC-accelerated systems, produced these estimates.


Compatibility With Existing Methods

SALMON is not a replacement for prior on-chip training algorithms. It is a structural addition that layers on top of existing methods, and this compatibility is one of its most useful properties for engineers working with established toolchains.

Tiki-Taka Gains

When Tiki-Taka, the dual-array asymmetry-compensation algorithm, combines with SALMON’s self-adaptive training, accuracy on ECRAM devices rises from 62.65% for Tiki-Taka alone to 70.96%, an improvement of 8.34%. On ReRAM devices, the combined approach reaches 38.90% versus 36.44% for Tiki-Taka alone. The gains vary across devices, but improvement appears consistently across all tested combinations, suggesting the two methods capture different types of error rather than overlapping on the same correction.

Mixed-Precision Gains

The mixed-precision method, which stores weight updates in FP32 digital memory rather than directly in the analog array, shows even larger gains from combination with SALMON. On ECRAM, mixed precision with self-adaptation reaches 92.13%, compared to 89.62% without it. On ReRAM, the combination produces 82.46% versus 73.06%, a difference of more than 9 percentage points. Mixed-precision methods rank among the more mature approaches to AIMC training and already have hardware support in several architectures, which makes this result directly relevant to practical deployment.


The Broader Implications

The SALMON work sits at a specific and important junction. Most research in neuromorphic and analog computing has focused on inference, taking a network trained elsewhere and running it on analog hardware. On-chip training has received far less attention. The nonidealities that make analog hardware fast and energy-lean also make it poorly suited to the precise arithmetic that gradient-based learning requires.

A Different Way to Think About Imperfection

The insight embedded in SAnet is that this tension does not have to resolve by making the analog hardware more precise. The network can instead internalize the hardware’s specific error profile. The digital attention blocks do not directly make the analog weights more accurate. They repeatedly inject a clean gradient signal that, over many training steps, reshapes the network’s internal representations. The analog weights end up in configurations that are robust to the hardware’s particular flavor of imprecision.

This is closer in spirit to how biological neural systems work than most conventional training algorithms. Biological neurons are noisy, variable, and far from precise. Yet biological systems learn to function reliably not by eliminating that variability but by organizing around it through experience. SALMON does not reproduce that process in any biological sense, but it arrives at a structurally similar solution through different means.

What This Means for Edge Deployment

For researchers and engineers working toward edge deployment of AI systems, energy budgets run in milliwatts rather than kilowatts. Training directly on analog hardware has been an important but elusive capability in that context. SALMON does not solve every aspect of that problem. The authors are clear that performance degrades under extreme nonideality conditions. However, for the device operating points that current and near-term analog memory technologies can realistically achieve, the self-adaptive approach produces both competitive accuracy and substantial energy reduction. That includes the ReRAM technologies available in commercial fabs today.

Open Questions

An open question is how often such systems would need to re-adapt as device characteristics drift over weeks or months of operation. Calibration drift after deployment, long-term conductance aging, and the question of retraining frequency on-chip all need further work before anyone can consider the approach fully production-ready.

The next steps for this line of research involve extending the approach to larger-scale networks and more complex datasets. Researchers should also explore whether analog attention blocks could replace the digital ones to reduce power further still, and map the warm-up and sparse strategies against real-world device aging profiles.

The path forward may not be to correct analog hardware, but to train systems that understand it.


Source: Young Woon Cho, Sungmin Lee, and Sangbum Kim, “SALMON: Self-Adaptive Learning Model on Neuromorphic Hardware,” IEEE Transactions on Neural Networks and Learning Systems, DOI: 10.1109/TNNLS.2026.3687785. The authors are with the Department of Materials Science and Engineering, the Inter-University Semiconductor Research Center, and the Research Institute of Advanced Materials, Seoul National University, Seoul, Republic of Korea.

Related Articles

PHP Code Snippets Powered By : XYZScripts.com
★

Track This Company

Create a free account to add companies to your watchlist and get notified when patents, funding, or key developments change.