Research

Advancing the frontiers of memory-reasoning integration through interdisciplinary research in neuroscience, mathematics, and robotics.

Research directions and objectives are shared here at a high level. Specific architectures, training procedures, and implementation details that form the VOLTAIC platform are proprietary to Verkko Robotics Ltd and protected under the Verkko Research Programme.

Verkko Research Areas

Memory-Reasoning Integration

Active Research

We're developing novel neural architectures where memory storage and reasoning processes are unified within individual nodes, inspired by how biological neurons integrate past experiences with current processing.

Research Objectives

  • Design memory manifolds that deform based on context and experience
  • Implement real-time memory-reasoning loops in hardware
  • Develop mathematical frameworks for adaptive memory structures
VOLTAICLiving Memory · Verkko Equilibrium Engine (VEE)

Biologically-Inspired Plasticity

Ongoing

Our research investigates how biological principles of neural plasticity can inform the design of adaptive systems that learn and evolve throughout their operational lifetime.

Research Objectives

  • Model synaptic adaptation mechanisms in artificial systems
  • Implement growth-based learning in spiking neural networks
  • Create environmentally responsive plasticity rules
VOLTAICSpiking Inference Core · Living Memory

Neuromorphic Hardware Implementation

Development

We're investigating hardware substrates for efficient implementation of adaptive spiking systems, focusing on event-driven processing and energy efficiency at the edge.

Research Objectives

  • Design execution pathways for sparse spiking workloads
  • Optimize event-driven processing architectures
  • Achieve ultra-low power consumption for edge deployment
VOLTAICSpiking Inference Core · Sovereign Edge Runtime

Embodied Cognitive Systems

Prototyping

Investigating how physical deployment environments shape the requirements for on-device learning — where the model must adapt continuously to conditions that cannot be anticipated at training time.

Research Objectives

  • Develop sensorimotor learning algorithms for continuous adaptation
  • Create adaptive control systems for dynamic environments
  • Demonstrate sustained real-world learning over extended horizons
VOLTAICSovereign Edge Runtime · Living Memory

Measured on neuromorphic silicon

Spike-native efficiency is not a projection. Published results on neuromorphic hardware already show inference in the microjoule to sub-millijoule range on keyword, gesture and vision workloads — an existence proof for the energy regime that event-driven architectures target.

Intel Loihi

0.24–0.31 µJ

per inference

0.8–1.3 ms latency on keyword and gesture workloads.

Intel Loihi 2

0.33–0.46 mJ

per inference

91–97% accuracy on vision-style classification tasks.

SpiNNaker 2

0.027 nJ

per spike

An energy floor for the spike event itself, independent of model.

BrainChip Akida

1.48 µJ

per inference

92% accuracy on keyword spotting in a commercial edge part.

Figures are reported by the respective hardware programmes and their published benchmark literature. They characterise the neuromorphic substrate rather than any Verkko system.

Open problems

Four questions define the neuromorphic frontier as we see it. None is solved, and we would rather publish the list than imply otherwise — they are also where collaboration is most useful to us.

01

Attention sparsification at scale

Carrying enforced sparsity through multi-head self-attention without the quality collapse that has defeated earlier attempts. This is the active step in the VOLTAIC roadmap.

02

On-device and online training

Learning in the field through ALIF surrogate gradients, inside the memory budget of a device that also has to keep running inference.

03

Cross-modal spike fusion at runtime

Fusing text, vision, audio and structured signals under one spike contract while they arrive, rather than aligning them offline.

04

Continuous learning in human–swarm context

Sustained adaptation across a population of agents sharing an environment with people, where what is worth remembering is not defined in advance.

Related Research Domains

Neuromorphic Computing

Brain-inspired computing architectures that process information using spikes and events

Loihi: A Neuromorphic Manycore Processor with On-Chip Learning

Davies et al. (2018)

IEEE Micro

Intel's Loihi chip demonstrates large-scale neuromorphic computing with on-chip learning capabilities.

Training Deep Spiking Neural Networks Using Backpropagation

Lee et al. (2016)

Frontiers in Neuroscience

Breakthrough in training SNNs using backpropagation through time, enabling deeper architectures.

Neuromorphic Electronic Circuits for Building Autonomous Cognitive Systems

Indiveri & Liu (2015)

Proceedings of the IEEE

Comprehensive review of neuromorphic circuits for autonomous cognitive applications.

Relevance to VOLTAICFoundation for our memory-reasoning integration approach using spiking neural networks

Synaptic Plasticity & Memory

How biological neural networks adapt and form memories through synaptic modifications

Synaptic Plasticity: Taming the Beast

Abbott & Nelson (2000)

Nature Neuroscience

Foundational work on understanding different forms of synaptic plasticity and their roles in learning.

The Self-Tuning Neuron: Synaptic Scaling of Excitatory Synapses

Turrigiano (2008)

Cell

Discovery of homeostatic plasticity mechanisms that maintain neural network stability.

Spike Timing-Dependent Plasticity: A Hebbian Learning Rule

Bi & Poo (1998)

Annual Review of Neuroscience

Seminal work establishing STDP as a fundamental learning mechanism in biological networks.

Relevance to VOLTAICBiological inspiration for our adaptive memory systems and plasticity mechanisms

Manifold Learning & Geometry

Mathematical frameworks for learning on non-Euclidean structures and manifolds

Geometric Deep Learning: Going Beyond Euclidean Data

Bronstein et al. (2017)

IEEE Signal Processing Magazine

Comprehensive framework for extending deep learning to non-Euclidean domains.

Representation Learning on Graphs: Methods and Applications

Hamilton et al. (2017)

IEEE Data Engineering Bulletin

Survey of graph neural networks and their applications to complex structured data.

Neural Ordinary Differential Equations

Chen et al. (2018)

NeurIPS

Continuous-depth neural networks that treat network depth as a continuous dimension.

Relevance to VOLTAICTheoretical foundation for memory manifold approaches and geometric neural architectures

Embodied AI & Robotics

How physical embodiment affects learning and intelligence in artificial systems

How the Body Shapes the Way We Think

Pfeifer & Bongard (2006)

MIT Press

Foundational book on embodied intelligence and morphological computation.

Developmental Robotics: From Babies to Robots

Lungarella et al. (2003)

Network: Computation in Neural Systems

Framework for understanding how robots can develop intelligence through interaction.

The Sensorimotor Approach to Perception

O'Regan & Noë (2001)

Behavioral and Brain Sciences

Theory that perception emerges from sensorimotor contingencies and active exploration.

Relevance to VOLTAICFramework for our embodied AI platforms and sensorimotor learning systems