Signed Weight Representation for Analog In-Memory Computing

28 concepts20 papers1 proposal

Analog In-Memory Computing (IMC) using crossbar arrays offers a massive efficiency advantage for deep learning by executing matrix-vector multiplications directly within memory substrates. By mapping inputs to voltages ($V$) and weights to physical conductances ($G$), the system leverages Ohm's Law ($I = G \cdot V$) and Kirchhoff's Current Law to accumulate products as currents along array columns.

However, this paradigm faces a fundamental physical limitation: conductance is strictly non-negative ($G \ge 0$). Conversely, artificial neural networks rely heavily on signed weights ($W \in \mathbb{R}$) to implement both excitatory (positive) and inhibitory (negative) synaptic connections. To bridge this gap, researchers have developed several hardware-software co-design strategies to represent signed values using unipolar analog components.

The Differential Pair Standard

The most prevalent solution to the sign problem is the use of a differential pair configuration (or dual-device mapping). Instead of mapping a single logical weight to a single memory cell, the weight is represented as the mathematical difference between two physical conductances:

$$W = G^+ - G^-$$

Here, $G^+$ represents the excitatory component and $G^-$ represents the inhibitory component.

        Input Voltage (V)
               │
       ┌───────┴───────┐
       ▼               ▼
   ┌───────┐       ┌───────┐
   │  G+   │       │  G-   │  (Physical Conductances)
   └───────┘       └───────┘
       │               │
       ▼ (I+)          ▼ (I-)
   ┌───────────────────────┐
   │ Differential Subtraction│ ──► Net Current (I_net = I+ - I-)
   └───────────────────────┘

This differential current-mode approach dates back to early neuromorphic designs, such as the amorphous-silicon memory arrays developed by A. J. Holmes et al. in Pulsestream Synapses with Non-Volatile Analogue Amorphous-Silicon Memories. Modern implementations apply this concept across various non-volatile memory technologies, including Resistive RAM (RRAM), Phase-Change Memory (PCM), and Charge-Trap-Flash (CTF). For instance, Varun Bhatt et al. utilized differential pairs of CTF cells in Software-Level Accuracy Using Stochastic Computing With Charge-Trap-Flash Based Weight Matrix to achieve high-precision deep learning inference.

Encoding Discrete and Ternary States

The differential framework easily accommodates quantized networks. In Ternary Neural Networks (TNNs), weights are restricted to $\{-1, 0, +1\}$. As demonstrated by Axel Laborieux et al. in Low Power In-Memory Implementation of Ternary Neural Networks with Resistive RAM-Based Synapse, these three states can be cleanly mapped onto binary-state RRAM cells:

  • Weight = +1: Excitatory cell in Low Resistance State (LRS), Inhibitory cell in High Resistance State (HRS).
  • Weight = -1: Excitatory cell in HRS, Inhibitory cell in LRS.
  • Weight = 0: Both cells in HRS (minimizing static power consumption).

By optimizing the sensing circuitry, the authors later showed in Implementation of Ternary Weights With Resistive RAM Using a Single Sense Operation Per Synapse that these ternary states can be read in a single clock cycle, preserving the throughput advantages of IMC.

Noise and Fault Mitigation

Beyond enabling signed arithmetic, differential mapping provides inherent robustness against hardware non-idealities. Stefano Ambrogio et al. demonstrated in An analog-AI chip for energy-efficient speech recognition and transcription that pairing PCM devices helps cancel out common-mode circuit offsets and low-frequency noise.

Furthermore, this redundancy can be exploited for fault tolerance. Geng Yuan et al. showed in Improving DNN Fault Tolerance using Weight Pruning and Differential Crossbar Mapping for ReRAM-based Edge AI that when a network is pruned, logical zeros can be intentionally mapped to align with physical "stuck-at-high-conductance" faults. If a device is broken and stuck in a high-conductance state, setting its differential partner to the exact same conductance masks the fault entirely, yielding a net weight of zero without requiring hardware replacement.

For more complex network architectures, such as LSTMs, the differential concept can be extended to a four-device bridge topology. Yi Liu et al. leveraged this configuration in Long-term and Short-term Memory Networks Based on Forgetting Memristors to manage the dynamic weight updates required for biological-style forgetting mechanisms.

Architectural Optimizations for Hardware Efficiency

While highly effective, standard differential mapping comes with a strict penalty: it doubles the physical area and the number of devices required, effectively cutting hardware utilization in half. To mitigate this overhead, researchers have introduced single-device and column-level alternatives.

Constant Offset Circuits

Instead of dedicating a second memristor column to every single weight column, a single crossbar array can be paired with a shared constant-offset circuit. As detailed by Haoqiang Guo et al. in Hardware Accelerator for Adversarial Attacks on Deep Learning Neural Networks, the physical conductances are biased such that a logical zero corresponds to a mid-range conductance value ($G_{mid}$). The net negative or positive weight is computed by subtracting a constant reference current at the edge of the array. This single-crossbar technique doubles hardware utilization compared to dual-crossbar designs.

Fine-Grained Polarization

Another approach is to group weights by sign mathematically before mapping them to hardware. In FORMS: Fine-grained Polarized ReRAM-based In-situ Computation for Mixed-signal DNN Accelerator, Yanzhi Wang et al. proposed a hardware-software co-design that enforces a sign constraint within fine-grained sub-arrays. By clustering weights so that all elements within a specific sub-array column share the same sign (either entirely positive or entirely negative), the architecture eliminates both the dual-device overhead and the need for complex analog offset circuits.

Standard Differential:  [ Col+ ][ Col- ] -> 2x Area Overhead
Polarized Sub-Arrays:   [ All+ ] or [ All- ] -> 1x Area (Sign handled at column boundary)

Input Voltage Inversion

An alternative perspective shifts the complexity from the weight representation to the input signal representation. Instead of manipulating the conductances to achieve subtraction, the system manipulates the input voltage polarity.

Michael Beigl demonstrated this principle in Realization and training of an inverter-based printed neuromorphic computing system using printed electronics. To represent a negative weight connection without negative resistance, the input voltage is passed through an inverter circuit to flip its polarity ($V \rightarrow -V$) before it enters the resistive crossbar. The resulting current contribution to the column accumulation is naturally subtractive:

$$I = G \cdot (-V) = - (G \cdot V)$$

This provides a low-overhead solution for analog multiply-accumulate operations, shifting the burden of signed arithmetic to minimal transistor logic at the input rows rather than duplicating memory cells across the entire array. Similarly, Runze Han et al. utilized a twin-cell architecture in A Novel Convolution Computing Paradigm Based on NOR Flash Array With High Computing Speed and Energy Efficiency where separate bit lines route positive and negative components to a differential subtractor at the output stage, enabling single-cycle signed 2D convolutions.

Summary of Trade-offs

Strategy Mechanism Key Advantage Primary Trade-off
Differential Pairs $W = G^+ - G^-$ High precision, noise/fault cancellation 2x hardware area overhead
Constant Offset $W = G - G_{mid}$ 1x device area utilization Requires precise analog reference circuits
Polarization Column-level sign constraints Eliminates hardware duplication Requires sign-aware training/clustering
Voltage Inversion Input polarity flipping ($-V$) Minimal transistor overhead Increases input routing/driver complexity

Key Foundational Work

Pulsestream Synapses with Non-Volatile Analogue Amorphous-Silicon Memories (1994) — contributed 1 concept elements to this direction.
Memristor crossbar based hardware realization of BSB recall function (2012) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_signed_weight_representation_57863e69,
  title = {signed weight representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/signed-weight-representation/rdr_57863e698fd64baf851771c4e32685a4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_signed_weight_representation_57863e69_v1,
  title = {signed weight representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/signed-weight-representation/rdr_57863e698fd64baf851771c4e32685a4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake