In-Situ Learning for Autonomous Neuromorphic Hardware

This direction is split into other directions: on-chip plasticity. Its last version stays readable and citable here.

Version 1 of 172 papers82 conceptsfrozen 2026-10-04

Most neuromorphic systems today suffer from an "inference gap." While they can execute pre-trained neural networks with extreme energy efficiency, they are effectively "frozen" once deployed. To update their weights, they must send data back to a power-hungry digital GPU, calculate gradients, and then reprogram the hardware.

The research direction of in-situ learning aims to close this loop by integrating training mechanisms directly into the physical substrate. By performing weight updates where the data lives—using the local physics of transistors, memristors, or optical circuits—these systems can achieve real-time adaptation with orders of magnitude less energy than traditional architectures.

The Bottleneck of Offline Training

Traditional computing follows the von Neumann architecture, where memory and processing are separate. In training, this creates a massive bottleneck: synaptic weights must be constantly moved between storage and the processor to be updated. As Conrad D. James and colleagues note in A historical survey of algorithms and hardware architectures for neural-inspired and neuromorphic computing applications, this reliance on external high-performance computers prevents neuromorphic chips from being truly autonomous.

In-situ learning eliminates this movement. Instead of treating the hardware as a passive container for weights, researchers like Kevin Portner and Till Zellweger demonstrate in Actor-Critic Networks with Analogue Memristors Mimicking Reward-Based Learning that weight update calculations can be performed directly within memristive crossbar arrays. This "in-memory training" allows the hardware to both store the synaptic state and compute the necessary adjustments simultaneously.

Leveraging Physics as a Gradient

The most intuitive leap in this field is moving from simulating math to harnessing physics. In a digital system, a gradient is a calculated numerical value. In an in-situ analog system, a gradient can be a physical current or a charge.

Paul E. Hasler and Carver Mead pioneered this with Single Transistor Learning Synapses, where the physical processes of electron tunneling and hot-electron injection are used to update weights "on the fly" during active operation. More recently, Ruibing Song and Ang Li introduced DS-LLM, which treats training as a quest for a minimum energy state in a dynamical system. By using physical currents to represent gradients, they achieve a 5,300x speedup in training because the weight updates are driven by the same energy already flowing through the system.

Algorithmic Adaptations for Hardware Constraints

Standard backpropagation is difficult to implement in hardware because it requires a global "backward pass" that is memory-intensive and biologically implausible. To solve this, researchers are developing "hardware-friendly" learning rules:

  1. Surrogate Gradients: Since spiking neurons have non-differentiable "jumps," Emre Neftci and Kenneth Stewart have used surrogate gradients to enable on-chip learning on processors like Intel’s Loihi. As shown in On-chip Few-shot Learning with Surrogate Gradient Descent on a Neuromorphic Processor, this allows chips to learn new gestures from just a few examples without needing a GPU.
  2. Forward-Only Learning: Adam A. Kohan and Hava T Siegelmann propose Forward Signal Propagation Learning, which eliminates the backward pass entirely. This simplification is ideal for neuromorphic hardware because it reduces the complexity of the required circuitry.
  3. Local Plasticity: Instead of global error signals, systems can use local rules like Spike-Timing-Dependent Plasticity (STDP) or the e-prop algorithm. Yujia Yin and colleagues demonstrate in Efficient Online Learning for Networks of Two-Compartment Spiking Neurons that these local updates allow for continuous learning with constant memory overhead, regardless of the sequence length.

Fusing Computation and Update

A key mechanism for efficiency is the Outer Product Accumulate (OPA) operation. In traditional training, calculating the weight update (the outer product of activations and errors) is a separate, serial step. Aayush Ankit and the team behind PANTHER showed that this can be fused directly into the memory crossbar. By performing the outer product as a single physical step, they eliminate the need to store intermediate gradients in off-chip memory, leading to a 103x reduction in energy.

The Path to Autonomous Edge Intelligence

The ultimate goal of this research is to move beyond static models. Whether it is Jung-Wook Cho using on-chip learning for Active Noise Canceling or Zhanglu Yan designing SparrowSNN for real-time ECG monitoring, the focus is on adaptation.

By integrating learning into the hardware, these systems can compensate for manufacturing defects, adapt to user-specific data patterns, and respond to changing environments—all while operating within the strict power budgets of wearable or embedded devices. As Thomas Ortner suggests in Learning-to-learn enables rapid learning with phase-change memory-based in-memory computing, the future of neuromorphic computing lies in hardware that isn't just "brain-inspired" in its structure, but "brain-like" in its ability to learn and evolve.

Papers in this version

Showing 60 of 72 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_in_situ_learning_for_autonomous_neuromor_22fe098a,
  title = {In-Situ Learning for Autonomous Neuromorphic Hardware},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/in-situ-learning-for-autonomous-neuromorphic-hardware/rdr_22fe098aa5d74f2e90fe671979cf2b10},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_in_situ_learning_for_autonomous_neuromor_22fe098a_v1,
  title = {In-Situ Learning for Autonomous Neuromorphic Hardware},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/in-situ-learning-for-autonomous-neuromorphic-hardware/rdr_22fe098aa5d74f2e90fe671979cf2b10/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}