Lab of Efficient Machine IntelligenceTU Delft

Algorithms ↔ Hardware

Efficiency, by design.

We study neuromorphic algorithm–hardware co-design for energy-efficient edge intelligence. Our work bridges learning methods, sensing systems and the hardware that makes them practical.

Explore the connections

One engine. Five directions.

Select a module to discover its research, experiments and publications.

THE EMI CO-DESIGN ENGINEEXPLORE THE LAYERS ↘
HARDWARE
ALGORITHMS
ALGORITHMS ↔ HARDWARESELECT A MODULE

Research direction 05

AI-Native Chip Design

AI agents compose, connect and refine circuit building blocks — from transistors to systems.

Agentic AI · chip design
AI agents compose an analog and digital circuitTransistor building blocks assemble into analog circuits alongside a grid of digital logic gates. Interconnects join the pieces into one system as the agent plans, composes, connects and refines the design. This is a concept illustration.DAI design agentANALOG CIRCUITSDIGITAL LOGICM3M4M1M2M5M6M7CcVBPVBNIN+IN−OUT
PlanComposeConnectRefine
Concept animation · analog + digital building blocks
ChipJevExplore our analog circuit design demo

Research direction 01

Learning better signals.

Learning-based models, calibration and digital predistortion for more efficient radio-frequency and mixed-signal systems.

With and without AI-DPD
With and without AI-DPDTwo synchronized paths compare the same radio signal and power amplifier. Without AI-DPD, the PA produces a compressed, distorted waveform. With AI-DPD before the PA, pre-correction restores a clean output. Both outputs remain visible together.Without AI-DPDDistortedDirectPAWith AI-DPDCleanAI-DPDPAWith and without AI-DPDTwo synchronized paths compare the same radio signal and power amplifier. Without AI-DPD, the PA produces a compressed, distorted waveform. With AI-DPD before the PA, pre-correction restores a clean output. Both outputs remain visible together.Same RF inputWithout AI-DPDDirectPADistortedWith AI-DPDAI-DPDPAClean
Same inputSame PACompare outputs
Same signal, same PA. AI-DPD pre-shapes the input to counteract distortion.
Explore signals
Digital predistortionRF power amplifiersADC calibration
OpenDPD RF measurement setup with a power amplifier, signal generator and spectrum analyzer.
AI for RF & mixed-signal systemsPhoto: OpenDPDv2 · Fig. 3
Ideas, in practice ↳2026OpenDPDv2: A Unified Learning and Optimization Framework for Neural Network Digital Predistortion2026Deep-Learning Based Hardware-Efficient Load Identification for Power Amplifiers

Research direction 02

Compute only what changes.

Brain-inspired algorithms and accelerators that exploit temporal and spatial sparsity to reduce unnecessary computation.

Dense vs. change-driven
Dense vs. change-drivenTwo matching networks receive the same four changed values, marked by orange rings. Dense computing activates all eighteen neurons. Change-driven computing activates only the four changed neurons, leaving fourteen crossed-out computations idle. Counts refer to this illustration, not measured performance.Same update · 4 changed valuesDense computing18 computationsChange-driven4 compute · 14 skippedOrange rings mark changes · × marks skipped workDense vs. change-drivenTwo matching networks receive the same four changed values, marked by orange rings. Dense computing activates all eighteen neurons. Change-driven computing activates only the four changed neurons, leaving fourteen crossed-out computations idle. Counts refer to this illustration, not measured performance.Dense computing18 computeChange-driven4 compute · 14 skip
Same changesCompareSkip repeats
Illustrative update: 4 of 18 values change. Dense computes all 18; change-driven computes 4 and skips 14.
Explore sparsity
Dynamic sparsityNeuromorphic computingNeural accelerators
The physical FPGA boards used for Spartus and Edge-Spartus on a wooden tabletop.
Sparse neural computingPhoto: Spartus · Fig. 10
Ideas, in practice ↳2025Exploiting neuro-inspired dynamic sparsity for energy-efficient intelligent perception2024Spartus: A 9.4 TOp/s FPGA-Based LSTM Accelerator Exploiting Spatio-Temporal Sparsity

Research direction 03

Perception at the edge.

Efficient visual perception, eye tracking and intelligent hardware for extended reality and wearable applications.

Seeing through events
Seeing through eventsAn eye looks from side to side. The moving edges produce event pixels, which guide a moving gaze marker.Eye motionEvent streamGaze estimateTrack the moving edges, without processing every pixelSeeing through eventsAn eye looks from side to side. The moving edges produce event pixels, which guide a moving gaze marker.Eye motionEventsGazeFollow changes. Update the gaze.
MotionEventsGaze
Eye motion becomes sparse events, then a continuously updated gaze estimate.
Explore vision
Event camerasEye trackingExtended reality
Participant using an eye-facing event camera beside a laptop during an eye-tracking demonstration.
Event-based vision & wearable intelligenceVideo: Q. Chen & C. Gao · 2023
Ideas, in practice ↳2026JaneEye: A 12-nm 2K-FPS 18.9-μJ/Frame Event-based Eye Tracking Accelerator2026BitFair: A 12nm Bit-Serial CNN Accelerator with Learnable Early Termination and Adaptive Bit Ordering for Ultra-Low-Power XR Vision

Research direction 04

Small devices. Useful intelligence.

Energy-efficient speech processing, keyword spotting and radar-based activity recognition for embedded systems.

Intelligence at the sensor
Intelligence at the sensorA microphone waveform and radar echoes flow into an edge AI chip. Its outputs identify voice and motion activity.SpeechRadarEdge AIVoiceMotionIntelligence at the sensorA microphone waveform and radar echoes flow into an edge AI chip. Its outputs identify voice and motion activity.SpeechRadarEdge AIVoiceMotion
Listen & senseEdge AIRespond
Speech and radar signals become useful information on a compact edge device.
Explore speech & sensing
Speech processingRadar sensingEmbedded intelligence
DeltaKWS keyword-spotting chip on its measurement board and MiniZed FPGA host.
Speech & intelligent sensingPhoto: DeltaKWS · Fig. 9(a)
Ideas, in practice ↳2026RadMamba: Efficient Human Activity Recognition through Radar-based Micro-Doppler-Oriented Mamba State-Space Model2025CleanUMamba: A Compact Mamba Network for Speech Denoising using Channel Pruning

EMI on GitHub

Open-source research.

Explore the code, datasets and models behind our work.

View all repositories
RF & mixed-signal AI

OpenDPD

An end-to-end learning framework for power amplifier modeling and digital predistortion, with measured datasets and reproducible benchmarks.

Code · Datasets · Benchmarks
Speech denoising

CleanUMamba

A compact Mamba network for speech denoising, with channel pruning, training tools and pretrained models for efficient audio processing.

Code · Pretrained models · Pruning
Radar sensing

AIRHAR

The official RadMamba implementation for radar-based human activity recognition, with data preparation, model training and evaluation workflows.

Code · Training · Evaluation

Research in action

From idea to demonstration.

Research support

Shared ambition.

Our research is supported by public funding, industry collaborations and a community of academic partners.

Grants, sponsors & awards
Chris and Yizhuo discussing our DPD demo