Ehsan Hosseini-Asl

I am a Senior Deep Learning Scientist at Nvidia AI , working on Large Language Models. Previously, I was a senior research scientist at the Salesforce AI Research. My research area is deep learning with application to natural language processing, speech processing and computer vision.

My current focus is building end-to-end language models for different NLP tasks. Specially, I am interested in building interactive language model which can make conversation with human and accomplish any NLP task. To that end, I proposed SimpleTOD, an end-to-end language model for task oriented dialogue (NeurIPS 2020 Spotlight).

I also work on few-shot learning, and sample efficient deep learning models that generalize better to different domains, including adversarial learning and robust speech recognition.

During PhD, I worked on building interpretable feature learning using non-negativity approach for neural network. This work extend to application in medical image analysis, for segmentation and accurate disease diagnosis models. The application includes lung segmentation, alzheimer's disease, prostate cancer, autism and renal rejection.

During my Master, I started working on neural network. I focused on building a new neural network based on wavelet transform , for learning irregular data, based on their sampling frequencies.

Email  /  CV  /  GitHub  /  Google Scholar  /  LinkedIn  /  Twitter

Research

I'm interested in machine learning, deep learning, natural language processing, dialogue, and computer vision.

Few-shot Aspect-based Sentiment Analysis with generative language model

Formulating Aspect-Based Sentiment analysis as autoregressive language modeling. It is a multitask model which achieves better few-shot performance with significant smaller standard deviation.

Joint Energy-based Model Training for Better Calibrated Natural Language Understanding Models

Using energy-based training (EBM) for fine-tuning language model on NLU tasks. The model reach better calibration with little to no loss in accuracy.

SimpleTOD: An end-to-end Language model for Task-Orinted Dialogue

Formulating Task-Oriented Dialogue as autoregressive language model. This way, the three subtasks, belief state prediction, action and response generation are modeled jointly.

Transferable multi-domain state generator for task-oriented dialogue systems

Proposing an encoder-decoder model for belief state generation in Task-Oriented Dialogue. The model is trained on multiple domains jointly, which achieves better few-shot performance using knowledge share across domains.

Augmented Cyclic Adversarial learning for low-resource domain adaptation

Augmenting Cyclic adversarial learning algorithm with a more stable cycle consitency loss using source model. This formulation stablize adversarial learning when number of labeled data in target domain is scarce.

Toward Scalable Neural Dialogue State Tracking Model

We proposed a global RNN encoder for belief state tracking across multiple slots. It improves inference latency by 35\% on average compared to the previous state-of-the-art model (GLAD)

A Multi-Discriminator CycleGAN for Unsupervised Non-Parallel Speech Domain Adaptation

GAN model training on spectrogram is unstable due to complexity in learning distribution of fine-grain frequencies. We propose a multi-frequency band discriminator into GAN model which learn to convert spectrogram from male to female.

PhD research

Research on representation learning in deep models. I worked on training a new autoencoder with non-negativity constraint which have better representation learning. This new autoencoder becomes the backbone of algorithms for medical image processing, including image segmentation and disease diagnosis.

Alzheimer's Disease Diagnostics by a Deeply Supervised Adaptable 3D Convolutional Network

Training a 3D convolutional neural network with deep supervision, improves Alzheimer classification accuracy.

Alzheimer's Disease Diagnosis by Adaptation of 3D Convolutional Network

Training a 3D convolutional neural network for Alzheimer classification. This network is pretrained with novel greedy layer-wise pretraining.

3D Lung Segmentation Using Incremental Constrained Nonnegative Matrix Factorization

An efficient segmentation algorithm based on our previous INMF approach for 3D lung ct scans. It works on voxels instead od pixel.

Automatic Segmentation of Pathological Lung Using Incremental Nonnegative Matrix Factorization

Our previous NMF-based model requires manual setting of the number of clusters for segmentation. Our new algorithm automatically detects number of clusters in the image, and in-homogeneities.

Deep Learning of Part-based Representation of Data Using Sparse Autoencoders with Nonnegativity Constraints

Learning part-based features in hidden layers of neural network, by constraining negative weights during training.

Lung Segmentation Based on Nonnegative Matrix Factorization

Our first algorithm for 3D lung segmentation using Non-negative Matrix Factorization. We treat each pixel and its 3D neighborhood as feature vector in NMF.

Other Medical Image Analysis

My phd work on nonnegative-constrained autoencoder (NCAE), results in building several efficient algorithms for image segmentation, feature extraction and disease diagnosis.

Artificial Intelligence (AI) Accurately Automate And Speed Immunofluorescence (IF)-based Discovery And Validation of Novel Prognostic And Predictive Biomarkers In Prostate Cancer

A New CAD System for Early Diagnosis of Autism Using Structural MRI

A comprehensive non-invasive framework for diagnosing prostate cancer

A New Non-Invasive Approach for Early Classification of Renal Rejection Types Using Diffusion-Weighted MRI

Computer-Aided Diagnosis Tool for Early Detection of Prostate Cancer

A New NMF-Autoencoder Based CAD System For Early Diagnosis of Prostate Cancer

MSc. research

Research on combining wavelet transform and neural network. The research results in proposing a neural natwork which can model regression data with different sampling frequencies.

Nonlinear dynamic system control using wavelet neural network based on sampling theory

Non uniform noisy data training using wavelet neural network based on sampling theory

Wavelet neural network based on sampling theory for non uniform noisy data

Design and source code from Jon Barron's website