Projects | Taliya Weinstein

Bias: Endangering Species and Performance

Sun, 26 Oct 2025 00:00:00 +0000

The project ‘Bias: Endangering Species and Model Performance’ investigates how long-tail imbalance and contextual bias in global wildlife classification models disadvantage endangered and data-deficient species. Using camera-trap data from Malawi’s Nkhotakota Reserve, Austin Kaburia and I evaluated SpeciesNet’s performance across species of differing IUCN conservation status. Results revealed a stark disparity in recall between common and vulnerable species (0.68 for impalas vs. 0.18 for leopards), exposing systematic underrepresentation of endangered taxa. To address this, the project fine-tuned SpeciesNet using taxonomic grouping, IUCN Red List metadata, and uncertainty quantification methods such as Generalized Additive Models (GAMs) and calibrated mixture models. These techniques improved endangered-species recall, introduced interpretable diagnostics for failure modes like nocturnality and foliage occlusion, and allowed the model to flag “unknown” cases for ranger review.

Beyond its technical contributions, the work aims to propose a new evaluation framework emphasizing endangered-species recall and calibration metrics aligned with conservation risk, filling a key gap in current literature. By bridging ecological insight with machine learning rigor, the project advances African-led AI for biodiversity monitoring and conservation safety. The work has been recognized with 3rd place at the 2025 African Computer Vision Summer School Hackathon, a Poster Award at the 2025 Deep Learning Indaba and a first place win at the Deep Learning Indaba Ideathon 2025, clearly underscoring the project’s scientific and societal impact.

A big thank you to Austin Kaburia for his research prowess and Kevin Kibaara for his digital media skills for our video submission.

Skills:

AI for Conservation, Bias Evaluation, Model Validation, Cross-Disciplinary Collaboration

Data Science for Health in Africa Workshop at Deep Learning Indaba 2025

Thu, 21 Aug 2025 00:00:00 +0000

Overview

Location

Kigali, Rwanda

The Data Science for Health in Africa (DS4Health) workshop took place on 21 August 2025 at the Deep Learning Indaba in Kigali, Rwanda. The event marked the fourth edition of the workshop and its largest turnout to date, bringing together over 100 researchers, healthcare professionals, and technologists from across Africa. Building on lessons from 2024, the workshop used insights from feedback reports to refine its structure, improve outreach, and strengthen engagement — resulting in a dynamic, high-impact event.

The day began with a keynote by Professor Michael Best (Georgia Institute of Technology), who spoke on designing technology with the recipient community in mind. This was followed by a panel discussion on The State of AI for Healthcare Research in Africa. As moderator, I designed the question set, guided the discussion, and ensured that perspectives from academia, industry, and clinical care were interwoven to highlight both progress and challenges in African health AI research.

Panelists included(from left): Prof. Michael Best (Georgia Institute of Technology) Dr. Rose Nakasi (Makerere University AI Health Lab) Dr. Arijit Patra (UCB Biopharma UK / University of Glasgow) Dr. Tobi Olatunji (Intron Health Founder) and Jay Patel (Jacaranda Health Chief Technical Officer)

The workshop was intentionally participatory, emphasising collaboration and peer learning. The Lightning Talk Competition showcased the work of emerging African researchers. As Programming Chair, I coordinated submissions and selected the winning presentations based on scientific merit and clarity.

Pictured here are the winners. The top scorer was Christopher Mvelase, followed by Sourour Ammar and Ferdaous Idlahcen.

The MedGemma tutorial, led by Dr. Mercy Asiedu and Dr. Sekou Remy from Google, provided hands-on experience in AI for medical imaging. The Design Dash followed, challenging interdisciplinary teams to design AI solutions addressing real healthcare challenges.

The day concluded with the launch of the Google-Sponsored DS4Health Ideathon. The Ideathon occured over 3 stages, with a final live pitch on the 9th of October 2025. I assisted with the Ideathon logistics and hosted the final live pitching event.

The 2025 workshop represented a significant leap in scale, visibility, and impact, much of which stemmed from using data and feedback gathered during the 2024 workshop to improve planning and delivery.

Key highlights included:

3.8× increase in attendance and 4.5× rise in feedback responses compared to 2024.
A clearer, more structured agenda, refined using insights from participant feedback forms.
Targeted advertising and pre-workshop seminars, which expanded awareness and attracted high-quality applications.
Improved activity flow, balancing expert talks with participatory sessions.
Consistently high satisfaction ratings, with 100% of respondents describing the workshop as “very useful.”
Strong praise for the Design Dash and panel discussion, which participants ranked as the most engaging and impactful sessions.

These outcomes validated the refinements made from the previous year. This demonstrates how structured evaluation and adaptive design can enhance workshop outcomes and build lasting research communities. The beginning section of the workshop is available for viewing.

From left in picture above: Ditiro Rampate, Charles Nimo, Comfort Adesina, myself, Ashery Mbilinyi and Marvellous Ajala

Roles and Responsibilities

I led both the conceptual and operational planning of the workshop and Ideathon. Key responsibilities included:

Organizing and moderating the AI in Health Research panel.
Selecting Lightning Talk winners and facilitating their presentations.
Designing the overall workshop structure and program flow, inclusive of co-authoring the proposal for the Deep Learning Indaba.
I managed on-ground logistics, and handled the liasing with main conference organisers.
Overseeing marketing, logistics, and budget management.
Coordinating pre-workshop seminars to build awareness and momentum.
Hosting and supporting the Ideathon.
Consolidating feedback and writing post-event outcome reports and blog features.

Skills:

Leadership, Technical Program Design, Event Planning, Panel Moderation, Workshop Facilitation, Budget Management, Marketing Strategy, Team Coordination, Research Communication , AI for Health Capacity Building

From Passive to Participatory: Healthcare Workshop Design for Continued Community Engagement

Sun, 27 Apr 2025 00:00:00 +0000

As part of my work with SisonkeBiotik, I led the writing of a conference paper for the HELINA 2025: Pan African Health Informatics Association. The paper examined key lessons learned from running workshops across different conferences, focusing on strategies for effective audience engagement and sustained community involvement. It argued that long-term impact requires workshops to go beyond one-off participation. This requires encourgaing collaborative idea formation among participants and creating spaces where these ideas can be developed and implemented. A poster based on this work was presented at the 2025 Deep Learning Indaba’s poster session and directly informed the design of the 2025 Data Science for Health in Africa Workshop.

Skills:

Workshop design, Writing, Data Collection, Data Analytics

Data Science for Healthcare Workshop at Deep Learning Indaba 2024

Sat, 07 Sep 2024 00:00:00 +0000

Overview

Location

U1 B12 University Amadou Makhtar Mbow, Rufisque, Senegal

The full day Data Science for Health workshop took place at the Deep Learning Indaba 2024 on the 6th of September. The workshop was organised with the intention of having casual exchnages of knowledge from external mentors as well as self-directed interaction acitvities to facilitate collaboration in future.

For the external mentors, we invited Dr Mercy Asiedu and MohammedElfatih MohamedKhair from Google to present on the cutting edge AI techniques being applied in health care and disaster response respectively. This was followed up with an insightful tutorial Gabriel Kallah-Dagadu to introduce participants to AI applied to healthcare data.

We then moved onto the interactive portion of the day which consisted of a Design Dash, whereby participants were tasked with brainstorming an AI solution to a current problem facing African healthcare systems. This activity was well recieved by participants with many noting how the dynamics of idea formation and subsequent presentation helped with future collaboration options. We then moved onto the lightning talk section, whereby 5 participnat finalists were given 5 minutes to present their research to the workshop audience. This was judged by Natasha Latysheva and myself. This was follwed with the finalists (chosen earlier in the week) of the ideathon presenting their solutions and the winner being announced.

The day concluded with an ethics panel with panelists Sibusisiwe Makhanya, senior research manager at IBM, Mwigereri Dorcas, AI in Health PhD researcher at Aga Khan University in Kenya, and Dr Sekou Remy, Seniour Research Scientist at IBM. The discussion ranged from the caution needed when deploying a model in a different location to where it was built, evaluating the need for the model at the problem level and the intracacies of applying for ethics when collecting or using secondary health data.

The workshop was enjoyed by the participants with the feedback forms reflecting that that attendees appreciated the mix of sessions and activites on offer. A big thank you to the in-person team of volunteers who made the workshop success a possibility! For a full list of organisers, see the website.

From left in picture above: Dr Francis Agamah, Daphne Machangara, Rolanda Julius, Comfort Adesina, Imane Hamzaoui , Chinonyelum Igwe, author and Gbetondji Donovan. Not picture Ditiro Rampate

Roles and Responsibilities

I played a key leadership role in organizing the workshop, defining subgroup tasks, actively engaging in meetings, and creating feedback forms to support future improvements.
I was responsible for helping to secure speakers, organizing the AI in Healthcare Ethics panel, and sourcing prizes. In the lead-up to the event, I managed on-ground logistics, and handled challenges like speaker cancellations and a last-minute venue change.
I facilitated the ethics panel and stepped in as a judge for the Lightning Talks.
After the event, I managed the accounting, fund distribution to prize winners, and compiled a detailed feedback report.

Skills:

Technical Communication, Event Planning, Ethics Panel Moderator, Fundraising, Lightning Talk Judge, STEM Education, Team Coordination, Workshop Creation

Embroidered Electronic Interfacing for Health Monitoring

Fri, 27 Oct 2023 00:00:00 +0000

A novel PEDOT: PSS-modified cotton thread (PECOTEX) was produced for sensor embroidery into T-shirts. This modified PECOTEX thread was interfaced with a commercial flexible printed board to monitor ECG signals from the clothing wearer. Originally, a T-shirt was utilised (as in the video). Thereafter, a sports bra was utilised for better skin contact. This project earned me the top Bioengineering student award for best thesis at Imperial College London.

Skills:

3D Printing, Analytical Problem Solving, Arduino, CAD design, Circuit Design, Data processing, Flexible Electronics, MATLAB, signal processing

Blood Pressure Group Statistics

Thu, 15 Dec 2022 00:00:00 +0000

In this MRes course project for the Computational and Statistical Methods course, blood pressure data for a placebo and treatment group is statistically analysed via Python in a Jupyter notebook. The core focus of the project was utilising the statistical analysis to arrive at conclusions and graphically represent the data. The project is self-contained with the data included for running the notebook.

Skills:

Data Analysis, Data Visualisation, Hypothesis Testing, Python, Statistics

Biomedical Unpacked Workshop Creation

Sun, 14 Aug 2022 00:00:00 +0000

Overview

This event introduced girls between the ages of 15 and 18 to the world of biomedical engineering through interactive presentations, group discussions with qualified engineers, video demonstrations, an in-depth Q&A session, and a hands-on microcontroller programming exercise to analyse photoplethysmography signals. The benefits of having attended this workshop included:

Learning about the engineering and biomedical engineering field to assess whether this is a viable career option
Gaining resources to plan a career in biomedical engineering
Interacting with biomedical engineering graduates and mentors
Developing technological skills in coding, sensor interfacing, and biomedical signals through a hands-on biomedical engineering activity

Roles and Responsibilities

Within my role as ambassador for the WomEng #1MillionGirlsinSTEM initiative, I created, planned, and presented a Biomedical Unpacked Workshop. The creation of this event included:

Designing a proposal for WomEng to introduce the benefits of this novel event.
Planning a whole day workshop inclusive of presentation content, hands-on coding activity, participant handout, volenteer hand outs and volunteer training.
Fundraising via the School of Electrical and Information Engineering at the University of the Witwatersrand for the requisite Arduino microcontrollers.
Installing the relevant software onto rented laptops.
Producing adverts, recruiting participants and producing application content.
Screening applicants and producing indemnity forms.

Skills:

Event Planning, Fundraising, Participant Recruiting, Photoplethysmography Reading, STEM Education, Workshop Creation

The Impact of Acoustic Features on the Persuasiveness of COVID-19 Lifestyle Speeches- Spectral and Cepstral Features

Wed, 27 Oct 2021 00:00:00 +0000

This project explored the persuasive impact of South African COVID-19 lifestyle speeches through a proof-of-concept investigation using eNCA interviews extracted from YouTube. Using generated spectral and cepstral features from interview audio, a neural network compares their impact on persuasiveness. The cepstral feature set outperforms the spectral set, with a balanced accuracy of 86.8% and an F1 of 85.0%, indicating its higher predictive power despite challenges like overfitting and dataset size. This project was completed with Gershon Koral.

Skills:

Database building, data processing, MATLAB, machine learning, signal processing

Weather Processing Statistics Project

Sat, 15 May 2021 00:00:00 +0000

In this course project for the final year class of Data Intensive Processing, algorithms to produce a 5-number summary of a large corpus of data in C++ and with multiple threads. The aim of this program is to analyse sets of daily precipitation data. The data is made available by the Global Precipitation Climatology Centre (GPCC). The program will output all the values necessary for a box-and-whisker plot. Processing has been implemented using MPI. The program accepts inputs for the dataset, the range of years to be analysed, and whether to include zeros in the analysis. The necessary error checks are performed on the input. This project was completed with Aaron Fainman and Nathan Jones.

Skills:

Computer Cluster, C++, MPI, Multi-threading, Parallel Computing, Statistics

Fundi Study Coordinator

Tue, 27 Apr 2021 00:00:00 +0000

This project was completed as part of my final year electrical engineering software development course and simulated the working of an agile workflow. Together with Basheq Tarifi, Nathan Jones, Yasser Karem, and Tarryn Maggs, we produced a full-stack software application for student study group coordination. The web application was deployed with Microsoft Azure and consisted of a NodeJS, Express, and Microsoft SQL database backend, and a front-end built with JavaScript, HTML, and CSS. Testing was done through frameworksJest for the backend and Cypress for front end-to-end testing.

Skills:

Agile development, CSS, Database design (SQL), Front end, HTML, Javascript, Microsoft Azure, NodeJS, User Testing

Dog Catcher Game

Sun, 27 Sep 2020 00:00:00 +0000

Designed and implemented a “pac-man” style game with the twist of dog and dog-catcher. The game made use of the SFML library in C++, with object-orientated design principles. This game won the second place prize for best software engineering project with a prize sponsored by Entelect. This project was completed with Aaron Fainman.

Skills:

C++, Game Design, Object Orientated Programming, Unit Testing

Steam Turbine Generator Simulator

Thu, 27 Apr 2017 00:00:00 +0000

This project was completed as part of my second-year electronics course with Alice Drozdov, Nicole Swanepoel and Ammaarah Akhalwaya. The project required each member to produce a different subsystem of the steam turbine generator simulation. The steam turbine generator is divided into subsections namely: the power supply, exhaust gas, steam valve, turbine shaft encoder, and condenser system. My subsystem was the steam valve system which utilised a relay, SN74S08N, AND logic gate, SN74LS74A D flip-flop for the processing and switching events.

Skills:

Hardware Design, Electronics, System Integration

WindAid Wind Turbine Construction

Sun, 01 Jan 2017 00:00:00 +0000

Overview:

WindAid is a non-profit organisation operating out of Trujillo, Peru to provide wind energy solutions to communities inaccessible to standard grid electricity, especially in underserved rural communities. WinAid attracts international students and professionals to assist with their programs which aim to have a new wind turbine built and installed over the month program.

Over my time at WindAid, I was involved in all facets of wind turbine production. This included the welding of the base support structure and its subsequent waterproofing spraying.

I was also responsible for assisting with producing the coils needed for the electromagnetic induction behind the electricity generation. Additionally, I was involved with assisting in the production of an anemometer system to monitor wind speeds for monitoring previously installed wind turbines.

The actual construction of the wind turbine blades was a multi-day event. It starts with lining the mold with fiberglass, then securing the two mold halves together and filling the mold with a foam mixture. The foam core is then removed and covered in fiberglass and carbon fiber and then a resin mixture is injected via a vacuum pump to saturate the mold fibers.

During my time at WindAid, the recipient community was a remote fishing village within Playa Blanca. The installation process included the wind turbine mounting, connecting the internal electrical wiring, and testing all connections.

Skills:

Anemometer, Arduino, Electrical Installation, Renewable Energy, Wind Turbine Creation