Algorithms

15 projects related to Machine Learning, Modeling, Datasets, and more!
Click on the tabs in the left to access the material of each presenter

Inducible Pluripotent Stem Cell (IPSC) Viability Prediction from Time-Lapse Microscopy Images

Researcher: Abhineet Singh | Supervisor: Nilanjan Ray 

The objective of this work is to develop a deep learning pipeline to make an early prediction of the viability of IPSCs to develop into insulin producing islet cells for type 1 diabetes patients. IPSCs have a minuscule viability rate of around 1 in 10,000 which makes it extremely costly to produce even a small number of usable cells since that requires culturing millions of cells to later stages of growth. The chemical assaying process needed to make this determination is also very tedious and labor-intensive. The ability to reliably predict which cells are likely to be viable at an early stage would therefore be ground-breaking in rendering this a practical and cost-effective treatment for type 1 diabetes.
Temporal information about the way that the appearance of a cell changes over time is crucial for predicting its future growth outcome. This data is collected by capturing time-lapse images of cell cultures from an ultra-high resolution microscope. These images are then manually annotated to mark the locations and identities of cells in order to train deep neural networks to perform automatic cell segmentation and tracking. Finally, cell classification is performed by multi-frame analysis of cell appearance through sequential deep networks.

 Pixel VQ-VAEs for Improved Pixel Art Representation

Researcher: Akash Saravanan | Supervisor: Matthew Guzdial 

Machine learning has had a great deal of success in image processing. However, the focus of this work has largely been on realistic images, ignoring more niche art styles such as pixel art. Additionally, many traditional machine learning models do not work well with pixel art due to its distinct properties. We propose the Pixel VQ-VAE, a specialized VQ-VAE model that learns representations of pixel art. We show that it outperforms other models in both the quality of embeddings as well as performance on downstream tasks. 

ADaPT-ML: A Data Programming Template for Machine Learning

Researcher: Andrea Whittaker | Supervisor: Denilson Barbosa 

Classification is a task that involves making a prediction about which class(es) a data point belongs to; this data point can be text, an image, audio, or can even be multimodal. This task can become intractable for many reasons, including:
• Insufficient training data to create a data-driven model; available training data may not be appropriate for the domain being studied, it may not be of the right type (e.g. only text but you want text and images), it may not have all of the categories you need, etc.
• Lack of available annotators with domain expertise, and/or resources such as time and money to label large amounts of data.
• Studying a phenomenon that changes rapidly, so what constitutes a class may change over time, making the available training data obsolete.
ADaPT-ML is a multimodal-ready MLOps system that covers the data processing, data labelling, model design, model training and optimization, and endpoint deployment, with the particular ability to adapt to classification tasks that have the aforementioned challenges. ADaPT-ML is designed to accomplish this by:
• Using Snorkel as the data programming framework to create large, annotated, multimodal datasets that can easily adapt to changing classification needs for training data-driven models.
• Integrating Label Studio for annotating multimodal data.
• Orchestrating the Labelling Function / Label Model / End Model development, testing, and monitoring using MLflow.
• Deploying all End Models using FastAPI.

Next Steps for Voyant and Spyral

Researcher: Bennett Tchoh | Supervisor:  Geoffrey Rockwell 

With more and more text available as electronic text, there is a need for more efficient ways to get insight from these texts. Voyant-tools (https://voyant-tools.org/) is a web-based set of tools for text analysis and visualization. It is open source and can also be downloaded and installed on windows and mac platforms. Although Voyant-tools is popular and widely used, its use is limited to the options available in the user interface. To address this limitation, we are developing a web-based JavaScript notebook environment called Spyral (https://voyant-tools.org/spyral) that builds on Voyant-tools. With Spyral we can call the tools used in Voyant and it allows for more complex and customized analysis. We can also comment, save and share our analysis. Spyral is still in its testing phased. We are currently developing tutorial notebooks to introduce new users to Spyral. This also permits us to identify and correct issues like bugs and glitches. Voyant and Spyral are constantly being updated and we are looking at the possibilities to include more functionalities so developers, contributors and testers are welcomed.. 

 Analyzing and Visualizing Twitter Conversations 

Researcher: Candelario A. Gutierrez Gutierrez | Supervisor: Eleni Stroulia

Social media platforms are public venues where users frequently participate in online conversations about topics of broad societal interest. For example, in Twitter they can interact in two ways: i) they can post an opinion or talk about an event, which is optionally accompanied with media, URL(s), hashtag(s) and/or handle(s); ii) they can like, share, reply and/or quote another user’s post. Private and public organizations have shown interest in this type of interactions to make event-driven decisions, and to achieve this, there is a necessity of Social Media Analytics (SMA) tools that facilitate the collection, processing, storage and extraction of insights from social media data. Hence, we present an automated and scalable data platform that contributes with three main components: a) an aspect based sentiment and a dictionary-based analysis within three lexicons, the personal values, sentiment, and humour; b) posts’ metadata and public’s engagement calculation; and c) feature rich visualizations to enable users to gain insights on their datasets. We demonstrate the usefulness of our platform with two case studies: (i) analyzing the fragmented narratives around established (hydro, oil and gas, coal, nuclear) and new (solar, wind, geothermal, biomass) energy sources; and (ii) comparing the social-media brands of academic institutions.

More at:

• Personal website: https://cande.me
• Project website: https://ai4buzz.ca

3D Human Behavior Modeling and Understanding

Researcher: Chuan Guo | Supervisor: Li Cheng

3D human behavior understanding and modeling aims to bridge the gap between 3D human behavior (e.g. motion, video) and other data domains such as action categories, control signals (e.g. directions, velocities, actions) and natural language, including topics of, for example, action recognition, motion captioning, or the inverse problems like human behavior simulation given actions, natural language descriptions or control signals.
In our project, we constructed large-scale datasets of annotated 3D human behaviors as well as developing a series of effective approaches to automatically understand, analyze and synthesize human behaviors in 3D environment.

Collaborative graph learning

Researcher: Fatemeh Tavakoli | Supervisor: Nidhi Hegde

Different organizations, companies, and even countries have invested a lot in creating or gathering valuable data for AI-developing purposes. Datasets owned by data owners may be biased, thus cannot generalize well over their interest population. For example, if we want to design drugs that are effective for all humankind, we need to share lab experiments on people of different nationalities. Or if we want to detect fraud, we need to have all the transactions in different financial institutions. For privacy and business reasons, it is not possible to share raw data. This research studies the problem of collaborative learning on graph-structured data. Consider many data owners that own a subset of a large graph. If we did not have any privacy concerns, the ideal case would be to connect these subgraphs according to their overlapping areas (shared nodes) to create a universal graph.
Our framework consists of a server and several data owners. Data owners encode their graphs into an embedding space using a GNN.
The goal of the server is to encourage the data owners’ models to encode their graphs into the same embedding space to reconstruct the universal graph in that embedding space and to enable collaborative training. 

Personalized and Explainable Aspect-based Recommendation using Latent Opinion Groups

Researcher: Maryam Mirzaei | Supervisors: Joerg Sander, Eleni Stroulia

The problem of explainable recommendation—lsupporting the recommendation of a product or service with an explanation of why the item is a good choice for the user—is attracting substantial research attention recently. Recommendations associated with an explanation of how the aspects of the chosen item may meet the needs and preferences of the user can improve the transparency and trustworthiness of consumer-oriented applications, which is the motivation driving this research area.
Current methods are far from ideal because they do not necessarily consider the following issues: (i) users’ opinions are influenced not only by individual aspects but also by the dependency between sentiments towards aspect; (ii) not all users place the same value on all aspects; and, (iii) any explanation are not provided for how the item aspects have led to the recommendation. We introduce a personalized explainable aspect-based recommendation method that can address these challenges. To identify the aspects that a user cares about, our semantics-aware method learns the likelihood of an aspect being mentioned in a user’s review. To capture dependency between the users’ sentiments towards an aspect, reviews that express opinions with similar polarities towards sets of aspects are clustered together in latent opinion groups. To construct aspect-based explanations, item aspects are rated according to their importance based on these latent opinion groups and the preferences of the target user. Finally, to provide a user with a (set of) useful recommendation(s) of an item, our method selects and synthesizes the aspects important for the target user.
We evaluate our method over two datasets from (a) Yelp and (b) Tripadvisor. Our results demonstrate that our method outperforms previous methods in both recommendation performance and explainability.

Collective Entity Linking with Quantum Annealing 

Researcher: Peiran Yao | Supervisor: Denilson Barbosa

Entity linking is an important step towards the understanding of language, as it bridges natural language processing with knowledge graphs. When there are multiple entity mentions within a piece of text, collectively linking these entities by considering their intercompatibility will improve the linking accuracy. However, collective entity linking is a computationally hard problem as it requires solving a non-trivial combinatorial optimization problem. Recently, we have witnessed rapid developments, and even commercialization, of quantum annealing, which leverages quantum effects to physically provide approximate solutions to many combinatorial optimization problems in an efficient way. In this work, we propose a quantum annealing algorithm for collective entity linking based on the D-Wave quantum annealer. We show that the algorithm can provide comparable accuracy and speed to traditional approximation algorithms like simulated annealing. We also discuss certain limitations of deploying this algorithm in real-world applications. 

Cluster for Analysis of Relational Data (CARD): A new research cluster for equitable data analysis

Researcher: Pete Jones | Supervisor: Deb Verhoeven

The Cluster for Analysis of Relational Data (CARD) is a new research lab set up at UofA to bring together researchers working in the area of social network analysis and relational data science. The explicit goal of the lab is to pioneer research which combines the computational and data-driven tools developed in these related fields with the intersectional feminist ethics of the lab’s participants. Such research, which explores research questions from the humanities and social sciences using computational methods, is becoming increasingly impactful due to the rapid technological advances underpinning such methods, which have made previously intractable questions about people and their social worlds suddenly much more accessible. However, AI and the algorithms which structure our digital world continue to be biased by the entrenched power dynamics and inequalities that structure the societies that create and imagine them. In this poster, we introduce CARD by providing an overview of some of the challenges posed by these biased algorithms, and open a discussion regarding the kinds of solutions that might enable algorithms to be a tool for good in computational social research. 

To Filter Prune, or to Layer Prune, That Is The Question

Researcher: Sara Elkerdawy | Supervisor: Nilanjan Ray 

Recent advances in pruning of neural networks have made it possible to remove a large number of filters or weights without any perceptible drop in accuracy. The number of parameters and that of FLOPs are usually the reported metrics to measure the quality of the pruned models. However, the gain in speed for these pruned models is often overlooked in the literature due to the complex nature of latency measurements. In this paper, we show the limitation of filter pruning methods in terms of latency reduction and propose LayerPrune framework. Layer- Prune presents a set of layer pruning methods based on different criteria that achieve higher latency reduction than filter pruning methods on similar accuracy. The advantage of layer pruning over filter pruning in terms of latency reduction is a result of the fact that the former is not constrained by the original model’s depth and thus allows for a larger range of latency reduction. For each filter pruning method we examined, we use the same filter importance criterion to calculate a per-layer importance score in one-shot. We then prune the least important layers and fine-tune the shallower model which obtains comparable or better accuracy than its filter-based pruning counterpart. This one-shot process allows to remove layers from single path networks like VGG before fine- tuning, unlike in iterative filter pruning, a minimum number of filters per layer is required to allow for data flow which constraint the search space. To the best of our knowledge, we are the first to examine the ef-fect of pruning methods on latency metric instead of FLOPs for multiple networks, datasets and hardware targets. LayerPrune also outperforms handcrafted architectures such as Shufflenet, MobileNet, MNASNet and ResNet18 by 7.3%, 4.6%, 2.8% and 0.5% respectively on similar latency budget on ImageNet dataset.

Online Search Under Convex Costs

Researcher: Siyuan Yu | Supervisor: Xiaoqi Tan

We propose online search under convex costs (OSCC), a new variant of online optimization problems that is fundamental to many applications such as electric vehicle charging, Internet advertising, financial trading, and various artificial intelligence and machine learning problems. In OSCC, agents arrive one at a time (e.g., sequential arrival of electric vehicles in a charging station), each with a value that remains unknown before its arrival. At each step when there is a new arrival, an irrevocable decision must be made in a “take-it-or-leave-it” manner, namely, to accept the agent and take its value, or reject it with no value. The crux of OSCC is that we must pay for an increasing and convex cost associated with the number of accepted agents – it is increasingly more difficult to accommodate additional agents. We are interested in social welfare maximization, namely, the sum of the accepted agents’ values minus the total convex cost.
We derive an adaptive threshold policy (ATP) that accepts an agent if its value exceeds an adaptively-changing threshold. Our ATP is logistically simple and intuitive to implement, and achieves the best-possible competitive ratio of all deterministic online algorithms. A key ingredient in our design is to connect the threshold function to the dual of the underlying offline optimization problem via the online primal-dual framework. We also give a lower bound on the competitive ratios of randomized algorithms, and prove that the competitive ratio of our ATP asymptotically converges to this lower bound when the number of agents that can be accommodated is sufficiently large. 

A Co-Simulation Environment for Learning Tasks over Power Grid and Communication Networks

Researcher: Talha Ibn Aziz | Supervisors: Ioanis Nikolaidis, Omid Ardakanian 

Generative Models Based on VAE and GAN for Data Synthesis

Researcher: Yuyang Nie | Supervisor: Anup Basu

In almost all supervised learning methods, the model’s quality depends on the input data. However, many research fields lack large-scale and diverse data sets. Image data collection and labelling can be challenging and costly in various tasks such as camera-capturing rare birds, labelling rare birds that requires domain expertise. Generative models like Generative Adversarial Networks (GAN) fill this gap by generating the data set needed by these models, and hence have a significant role in future AI applications.
In our work, we developed a generative model using Variational Autoencoder (VAE) and GAN to synthesize new image data. We have implemented a VAE model and used it as the generator for GAN and train the model on the “Caltech-UCSD Birds 200” dataset. The novelty about our work is that we added an extra component in the model to incorporate bird types. As a result, the model can not only generate more diverse images within the same category, but also counter-factual image data (e.g. sparrow-looking crow).

Generating Private Time Series Dataset

Researcher: Zahra Bashir | Supervisor: Nidhi Hegde 

Some datasets contain sensitive private information that should not be shared publicly. The necessity of having a synthetic dataset that is so close to the original one but not leaking any private information appears here. I’m working on private data generation using Time-GANs for time series. Time-GAN is a natural framework for generating realistic time-series data in different domains. The model contains an Embedder, a Recovery, a Generator, and a Discriminator. For instance, I’m trying to generate an activities dataset(for one subject) in which we cannot tell that person A did X at time T. This is what I define as a private synthetic dataset. Differential Privacy(DP) and Pufferfish Privacy(PP) happen to be some concepts I’m interested in. I’m trying to preserve Time-GAN using DP and PP in terms of privacy. I’ll apply DP during the Time-GAN training to add randomness to the gradients(DPSGD), leading to a more private model that cannot be attacked(for instance, inference attack). Also, I’ll apply a mechanism named Markov Quilt(in the PP category) to the synthetic data directly to make the data private. I’m going to compare these two approaches to see which is more helpful in terms of fidelity, privacy, and utility.

AI and Social Media Identity

Project description:

Supervisor:

Description:
Supervisor:

sdafsdf