RESEARCH INDEX

The research record

10SELECTED WORKS · 2017 → 2026
FEATURED
HOMINIS field
HOMINIS — a human-centered foundation model for Europe
10 JUL 2026 · PREPRINT · DKAI LAB × REALAI
RESEARCH AREAS — 04
01
Human-centered foundation models
02
Subsurface & geoscience AI
03
Learning rules & activations
04
Applied imaging
FIGURE STUDY — 2019 · OPEN SOURCE

Dynamic Linear Flow

A family of invertible transformations with a partially autoregressive structure — state-of-the-art among flow-based models on ImageNet 32×32 and 64×64, converging roughly ten times faster than Glow, with the code released. Research by Huandong Liao at the lab.

The Dynamic Linear Flow architecture: squeeze, invertible 1×1 convolution, dynamic linear transformation and split
FIG. 1 — THE DLF BLOCK: SQUEEZE → INVERTIBLE 1×1 CONV → DYNAMIC LINEAR TRANSFORMATION → SPLIT
Linear interpolation through the model's latent space between two unseen images
FIG. 2 — LATENT-SPACE INTERPOLATION BETWEEN UNSEEN IMAGES
Samples generated by the model, trained on CelebA-HQ
FIG. 3 — MODEL SAMPLES, CELEBA-HQ 256×256
ALL PUBLICATIONS
HOMINIS — a human-centered foundation model
2026 · PREPRINT
VeerNet — seismic interpretation under domain shift
2023 · PEER-REVIEWED
EarthAdaptNet — subsurface domain adaptation
2020 · PEER-REVIEWED
DentAI — AI for dental image analysis
2020 · ITU / WHO AI4H
CytoDeep — deep vision at microscopy scale
2019 · APPLIED
Dynamic Linear Flow (DLF) — generative model
2019 · OPEN SOURCE · HUANDONG LIAO
Neuroscience-inspired learning rules for deep networks
2019 · WORKSHOP
ARiA — an adaptive activation family
2018 · PREPRINT
Capsule routing in practice — engineering notes
2018 · NOTES
Capsule Networks in TensorFlow
2017 · OPEN SOURCE
OPEN SOURCE — THE 2017 LEGEND

Days after the paper, the world ran capsules on our code.

github.com / deepkapha
NOV 2017 · PYTHON · TF 1.X
View on GitHub ↗