KNOWLEDGE
INVESTMENT GROUP

SELECTED RESEARCH · 2015–2023

Research,
made useful.

Ten distinct works coauthored by Andrei Ionut Damian and collaborators. Explore the questions, methods and publications behind research in predictive analytics, computer vision and AI systems.

This collection covers 2015–2023; the verified papers below were first published from 2017 onward. Conference papers and preprints are identified individually, with versions of the same work grouped together.

Andrei Ionut Damian on Google Scholar
BROWSE BY YEAR202320212020201920182017

20231 work

PreprintDistributed AI & MLOps

AiXpand AI OS — Decentralized ubiquitous computing MLOps execution engine

Beatrice Milik · Stefan Saraev · Cristian Bleotiu · Radu Lupaescu · Bogdan Hobeanu · Andrei Ionut Damian

arXiv:2306.08708v1 · 14 June 2023

Explores low-code AI applications distributed across a cooperative network of computing resources. The proposed system builds on the SOLIS framework and addresses job allocation, communication and decentralized execution through a shared-resource architecture.

  • Cooperative deployment of end-to-end AI pipelines
  • Secure job distribution and shared computing resources
Read the 2023 preprint

Original 2023 version and author list. Later revisions use a different title and are outside this collection.

20212 works

PreprintDistributed AI & MLOps

SOLIS — The MLOps journey from data acquisition to actionable insights

Razvan Ciobanu · Alexandru Purdila · Laurentiu Piciu · Andrei Damian

arXiv:2112.11925 · 2021, revised January 2022

Addresses the work between a successful model experiment and an operating AI application. SOLIS connects data acquisition, parallel inference, business logic and communications in a configurable pipeline that can run on edge devices or cloud machines.

  • Plugins for capture, serving, business logic and communication
  • Integration with multiple model frameworks and messaging protocols
Read the preprint · 2022 revision

First submitted in December 2021; the linked revision is from January 2022.

Conference paperRetail & predictive analytics

ProVe — Self-supervised pipeline for automated product replacement and cold-starting based on neural language models

Andrei Ionut Damian · Laurentiu Piciu · Cosmin Marinescu · Nicolae Tapus

CSCS 2021 · pp. 98–105

Applies ideas from language representation to retail transactions. Product vectors help identify alternatives to unavailable items and introduce products with little sales history. The pipeline combines learned relationships with product information to support replacement, category assignment and demand forecasting.

  • Product replacement and complementary-item relationships
  • Cold-start handling for newly introduced products
Conference paper · 2021 Preprint · 2020, revised 2021 Original 2020 submission

One work, two publication forms: a 2020 preprint and a 2021 conference paper. The preprint lists Andrei Ionut Damian, Laurentiu Piciu and Cosmin Mihai Marinescu; the conference version also credits Nicolae Tapus.

20201 work

Conference paperNeural architectures

A view on automated neural graph topology generation and a viable direction of innovation

Andrei Ionut Damian · Laurentiu Piciu · Nicolae Tapus

RoEduNet 2020 · pp. 1–7

Reviews approaches to automatically designing neural-network structures. The work proposes MultiGatedUnit, an experimental architecture using learnable self-gating mechanisms, as a direction for reducing the manual search involved in choosing a network topology.

  • Review of neural architecture generation
  • Experimental learnable gating through MultiGatedUnit
Conference paper

MultiGatedUnit is presented as an experimental research direction.

20192 works

Proceedings articleComputer vision & cloud software

CloudifierNet — Deep Vision Models for Artificial Image Processing

Andrei Ionut Damian · Laurentiu Piciu · Alexandru Purdila · Nicolae Tapus

Procedia Computer Science 162 · pp. 720–728 · ITQM 2019

Studies a different kind of visual scene: software interfaces and hand-drawn screen mockups. The research develops convolutional models and purpose-built training data for recognizing interface elements, contributing a visual-analysis step toward automated software understanding and migration.

  • Artificial interface scenes and hand-drawn mockups
  • Purpose-built vision models compared with transfer-learning approaches
Published article Read the 2019 preprint

The experiments address visual analysis within a broader research programme on software understanding and migration.

Conference paperRetail & predictive analytics

Advanced Customer Activity Prediction Based on Deep Hierarchic Encoder-Decoders

Andrei Damian · Laurentiu Piciu · Sergiu Turlea · Nicolae Tapus

CSCS 2019 · pp. 403–409

Models purchasing activity as a sequence of baskets rather than isolated product choices. Hierarchical encoder-decoder models capture both the contents of a shopping session and patterns across sessions, supporting next-basket recommendations and behavioral segmentation.

  • Representations of individual baskets and longer customer journeys
  • Next-basket experiments reported on the Ta-Feng retail dataset
Conference paper Read the 2019 preprint

The paper reports next-basket experiments and discusses further work on jointly learning the customer journey and recommendation sequence.

20182 works

Conference paperRetail & predictive analytics

Deep Neural Pipeline for Churn Prediction

Andrei Simion-Constantinescu · Andrei Ionut Damian · Nicolae Tapus · Laurentiu-Gheorghe Piciu · Alexandru Purdila · Bogdan Dumitrescu

RoEduNet 2018 · pp. 1–7

Investigates customer retention through deep neural models and time-to-next-event prediction. The pipeline uses GPU computation to work with transactional data, addressing both the likelihood of a customer leaving and the timing of their next activity.

  • Neural models for customer churn
  • Time-to-next-event prediction and parallel computation
Conference paper · 2018 Author’s introduction · 2019

The authors’ 2019 article provides an accessible introduction to the 2018 conference publication.

Conference paperRetail & predictive analytics

Deep recommender engine based on efficient product embeddings neural pipeline

Laurentiu Piciu · Andrei Damian · Nicolae Tapus · Andrei Simion-Constantinescu · Bogdan Dumitrescu

RoEduNet 2018 · pp. 1–6

Learns product and customer representations from sequences of retail transactions. The resulting vectors support product similarity, complementary-item recommendations and sales prediction. Experiments use pharmaceutical retail data, with unsupervised representation learning followed by supervised prediction.

  • Joint product and customer embedding spaces
  • A recommendation and sales-prediction pipeline evaluated on retail data
Conference paper · 2018 Read the 2019 preprint

The conference paper is from 2018; its openly accessible arXiv version was deposited in 2019.

20172 works

Conference paperComputer vision & cloud software

Cloudifier virtual apps: Virtual desktop predictive analytics apps environment based on GPU computing framework

Andrei Ionut Damian · Alexandru Purdila · Nicolae Tapus

IEEE ICCP 2017 · pp. 133–138

Presents the architecture and early experimentation of an online environment for hosting and generating applications with predictive capabilities. GPU-based parallel computing provides the foundation for bringing inferential analysis into a virtual application environment.

  • A platform architecture for predictive applications
  • GPU computing as an application-building resource
Conference paper · 2017 Author’s introduction · 2019

The author’s explanatory blog post appeared in 2019; the conference publication dates to 2017.

Conference paperComputer vision & cloud software

Model Architecture for Automatic Translation and Migration of Legacy Applications to Cloud Computing Environments

Andrei Ionut Damian · Nicolae Tapus

CSCS 2017 · pp. 577–582

Proposes a pipeline architecture for translating legacy applications into cloud-enabled environments. The motivating challenge is the large body of desktop and client-server software whose migration remains costly, particularly for organizations unable to fund a complete redevelopment.

  • An architecture for automated application translation
  • A research approach to reducing migration effort
Conference paper · 2017 Author’s introduction · 2019

The publication presents the proposed architecture and its motivation.

FROM THE TEAM’S NOTEBOOK

Beyond the paper.

Historical writing, prototypes and learning initiatives from Lummetry.AI.

2020 · TEAM NOTE

Sharing practical data-science experience

Lummetry.AI announced hands-on online deep-learning and data-science classes using Microsoft Azure, with support from Microsoft.

Read the original post
2019 · TEAM NOTE

Working with less data

Andrei’s essay discusses few-shot, one-shot and zero-shot learning, synthetic training data and the privacy questions that arise when generating data.

Read the original post
2020 · TEAM NOTE

An embedded AI prototype

The OmniDJ prototype announcement describes a modular IoT system applying computer vision to audience activity in hospitality venues.

Read the original post

THE NEXT CHAPTER STARTS WITH A CONVERSATION

What could we
build together?

Let’s explore the possibilities