Research

GCVAE: Generalized-Controllable Variational Autoencoder

Open in Colab

Authors: Kenneth Ezukwoke, Anis Hoayek, Mireille Batton-Hubert, Xavier Boucher

GCVAE is a generalized version of VAE with adaptive weights. It solves the problem of disentanglement and reconstuction trade-off by replacing fixed hyperparameters with trainable Lagrangian hyperparameters:

  • Maximizing mutual information \(I_p(x, z)\) during data reconstruction (\(x\): data; \(z\): latent)
  • Data reconstruction is done \(w.r.t\) inference constraints (\(e.g.\), \(Kullback–Leibler\) \(divergence\))
  • Optimization framework designed to introduce controllable and adaptive Lagrangian hyperparameters
  • Proportional-Integral-Derivative (PID) controller serves for automatic controlling of hyperparameters (\(\alpha, \beta, \gamma\))

The controllable hyperparameters ensures simultaneously learning a well disentangled representation with high reconstruction quality and minimal information loss during the bottleneck-compression.

Leveraging Pre-trained Models for Failure Analysis Triplets Generation

Authors: Kenneth Ezukwoke, Anis Hoayek, Mireille Batton-Hubert, Xavier Boucher, Pascal Gounet, Jerome Adrian

The research focuses on leveraging pretrained language models (such as transformers) for failure analysis (diagnosis) in semiconductor industry.

  • Generative Pretrained Transformers (\(GPT2\)) is best for the task of Failure Analysis Triplet Generation (FATG)
  • Sequential structured data is challenging for Encoder-Decoder Transformers especialy for data-to-text problems
  • BLEU, ROUGE and METEOR scores do not correlate with human evaluation when scoring sequential data-to-text problems
  • We introduce Levenshstein Sequential Evaluation (\(LESE-N\)) metric, which correlates with human-evaluation

Significantly improved results fine-tuning pretrained GPT2 (a decoder only transformer) for the downstream task of FATG. Next, we introduce the GCVAE variational loss for fine-tuning.

Big GCVAE

Authors: Kenneth Ezukwoke, Anis Hoayek, Mireille Batton-Hubert, Xavier Boucher, Pascal Gounet, Jerome Adrian

This work is the convergence resulting from finetuning Large Language Model (LLM) with the Generalized-Controllable Variational AutoEncoder (GCVAE), for improved Failure Analysis Triplet Generation (FATG)

  • Introduces latent space modeling for domain generalization, hence improving inference and generation
  • Finetuned on Generalized-Controllable Variational AutoEncoder loss (GCVAE) exhibits superior performance over stateof-the-arts
  • Demonstrates capacity to generate failure analysis triplets specifically tailored to root cause identification in product packages
  • Self-supervised model that operates with adaptive-controllable hyperparameters, eliminating the need for human intervention in the decision-making process

An improved domain specific (FATG) LLM model that introduces latent modeling for domain generalization and generation.

FORECASTING-1.0 / Automated Signal Generator

Author: Kenneth Ezukwoke

A financial research software with the capacity to automatically generate signals for algorithmic trading using OANDA v20 REST API as data source.

  • Data pool request for currencies, metals, derivatives and commodities for different timeframes (e.g., H30, H6, H8, H12, W) based on OANDA v20 REST API
  • Built on well known trend and price-based technical indicators (e.g., SMA, EMA, MACD, HMA, Stoch-Oscilator, CCI, SuperTrend) for signal generation
  • Interactive Graphical User Interface (GUI) for seemless strategy control

Disclaimer: FORECASTING-1.0 and ASG are research tools for educational and experimental purpose only. Software is subject to users discretion and open to financial risk upon usage.

Selected projects

Project Description Git
Advance Machine Learning- Kernel Methods Implementation of key advance ML algorithms (both classical and kernel versions)
Transfer Learning and Optimal Transport Subspace Alignment & Sinkhorn’s Algorithm
Social Mining Recommendation System Content based Recommedation system based on Graph mining
Active and Online Learning Machine Learning Algorithms Online learning and Active learning ML algorithms
FORECASTING 1.0 An intelligent model for time-series prediction and forecasting
Unsupervised Anomaly Detection Unsupervised machine learning methods for novel anomaly detection
NLP Pproject Book Insights with Plotly Machine learning for Natural Language Processing (NLP)
Stock Return Prediction using KNN-SVM-GAUSSIAN-ADABOOST Forecast stock prices using classical machine learning techniques