Postdoctoral Research Fellow in Artificial Intelligence

Jesse Wood

I build deep learning models that read the chemical fingerprints of fish, to authenticate seafood and help protect the ocean it comes from.

Te Herenga Waka—Victoria University of Wellington

About

I am a postdoctoral research fellow in the School of Engineering and Computer Science at Te Herenga Waka—Victoria University of Wellington, working with Hiringa te Mahara, the Aotearoa New Zealand pod of Abundant Intelligences, an Indigenous-led programme that reimagines AI from Indigenous knowledge systems. My postdoc is supervised by Dr Kevin Shedlock. I collaborate with the Bioeconomy Science Institute, formerly Plant & Food Research, and sit on the board of the AI Research Association of Aotearoa New Zealand as the elected representative for postdoctoral researchers.

I hold a BE (Hons) in Software Engineering and a PhD in Artificial Intelligence. Before the PhD I was a software developer intern and later a contract software engineer at NIWA, New Zealand's National Institute of Water and Atmospheric Research. I still think of code as a scientific instrument: something to build carefully, test and then point at an open question. Off the clock I built a Skyrim-style wellbeing tracker and am currently lost somewhere in Baldur's Gate 3.

Research

My research sits where machine learning meets analytical chemistry. A mass spectrometer or a gas chromatograph turns a piece of fish into a long vector of numbers, and the interesting question is what a model can learn from that signal: which species it is, which part of the fish, whether it has been adulterated and which batch it came from. Right now that means deep learning on GC-MS chromatograms to authenticate fish species, framed through kaitiakitanga and aimed at protecting marine resources from illegal, unreported and unregulated fishing. My thesis did the same for rapid evaporative ionisation mass spectrometry, where transformers reached 100% accuracy on species identification and a self-supervised method, SpectroSim, traced samples to their batch without labels. I care most about self-supervised learning, transformer architectures and evaluation rigorous enough that a chemist would trust the result.

Publications

  1. · PhD thesis · Te Herenga Waka—Victoria University of Wellington

    Machine Learning for Rapid Evaporative Ionization Mass Spectrometry for Marine Biomass Analysis

    Jesse Wood

    Abstract

    This thesis advances seafood processing by applying deep learning to Rapid Evaporative Ionization Mass Spectrometry (REIMS) data, enabling automated and accurate marine biomass analysis.

    This research addresses critical industry challenges in quality control, food safety, and fraud prevention. These include foundational tasks like species identification to combat mislabeling fraud and body part classification to optimize by-product utilization and prevent adulteration with offal. It also formalizes novel food safety problems, such as detecting oil contamination (a food safety hazard) and cross-species adulteration (a form of economic fraud). Finally, it tackles batch traceability, a task essential for rapid product recalls that currently relies on costly and impractical physical tagging methods.

    To achieve this, the thesis first establishes a suite of advanced models for foundational tasks species and body part identification, drawing on datasets described in Chapter 3. Specifically, this involves a binary classification dataset of 106 samples for Hoki and Mackerel identification and a multi-class dataset of 33 samples for classifying seven distinct body parts (fillets, heads, livers, skins, gonads, guts, and frames). The work then formalizes and solves novel food safety problems, including oil contamination (seven ordinal concentration levels from 50% to 0% in a 126-sample dataset) and cross-species adulteration (pure Hoki, pure Mackerel, and mixed Hoki-Mackerel samples from a 144-sample dataset), also based on Chapter 3 data. Finally, it introduces a self-supervised framework to address the challenge of batch traceability using a highly imbalanced, pairwise comparison dataset of 2,556 instances, derived from 72 fish samples across 24 distinct batches. The high-dimensional nature of the REIMS data, with 2,080 features per spectrum for all tasks, necessitates sophisticated solutions.

    Key contributions include the development of advanced Transformer architectures, which significantly outperform traditional methods by achieving up to 100% accuracy in species identification. The study also critically evaluates Mixture of Experts (MoE) Transformers to determine optimal configurations and investigates the asymmetric effects of transfer learning across different tasks, namely species identification, body part classification, oil contamination, and cross-species adulteration detection. Furthermore, a comparative analysis of ordinal classification techniques is included, demonstrating that models designed for ordered data significantly reduce error distance in graded contamination tasks. For batch traceability, a novel self-supervised contrastive learning method, SpectroSim, is introduced, which identifies sample origins with 70.8% accuracy without using class labels. Explainable AI techniques are employed to ensure model predictions are interpretable, enhancing trust and providing chemically relevant insights.

    This research is motivated by the need to overcome the limitations of current state-of-the-art analytical methods. The standard approach for REIMS analysis relies on traditional chemometric techniques like Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA), which are often limited in their ability to model the complex, non-linear, and high-dimensional nature of mass spectra. For the foundational task of species and body part identification, the challenge was to classify samples from complex and high-dimensional REIMS data. This was addressed by developing novel Transformer architectures that achieved near-perfect accuracy (up to 100%), establishing a new performance benchmark. For the novel tasks of oil contamination and cross-species adulteration, the challenge was to detect subtle chemical signatures often masked by the sample’s primary chemical profile. This was solved by adapting advanced models to these difficult classification problems and systematically applying transfer learning, which was found to consistently improve the detection of oil contamination. Finally, for batch traceability, the challenge was to create a method that avoids costly physical tags and the need for impractical supervised labels. This was achieved by developing “SpectroSim”, a self-supervised framework that accurately determines sample origin without labels, offering a practical and cost-effective solution for the industry.

    BibTeX
    @phdthesis{wood2026machine,
      title={Machine Learning for Rapid Evaporative Ionization Mass Spectrometry for Marine Biomass Analysis},
      author={Wood, Jesse},
      school={Te Herenga Waka---Victoria University of Wellington},
      year={2026},
      doi={10.26686/mef0-mr46}
    }
  2. · Journal article · Intelligent Marine Technology and Systems

    Hook, Line and Spectra: Machine Learning for Fish Species and Part Classification using Rapid Evaporative Ionization Mass Spectrometry

    Jesse Wood, Bach Nguyen, Bing Xue, Mengjie Zhang, Daniel Killeen

    Abstract

    Marine biomass composition analysis traditionally requires time-consuming processes and domain expertise. This study demonstrates the effectiveness of rapid evaporative ionization mass spectrometry (REIMS) combined with advanced machine learning (ML) techniques for accurate marine biomass composition determination. Using fish species and body parts as model systems representing diverse biochemical profiles, we investigate various ML methods, including unsupervised pretraining strategies for transformers. The deep learning approaches consistently outperformed traditional machine learning across all tasks. For fish species classification, the pretrained transformer achieved 99.62% accuracy, and for fish body parts classification, the transformer achieved 84.06% accuracy. We further explored the explainability of the best-performing and predominantly black box models using local interpretable model-agnostic explanations and gradient-weighted class activation mapping to identify the important features driving the decisions behind each of the best performing classifiers. REIMS analysis with ML can be an accurate and potentially explainable technique for automated marine biomass composition analysis. Thus, REIMS analysis with ML has potential applications in quality control, product optimization, and food safety monitoring in marine-based industries.

    BibTeX
    @article{wood2025hook,
      title={Hook, line, and spectra: machine learning for fish species identification and body part classification using rapid evaporative ionization mass spectrometry},
      author={Wood, Jesse and Nguyen, Bach and Xue, Bing and Zhang, Mengjie and Killeen, Daniel},
      journal={Intelligent Marine Technology and Systems},
      volume={3},
      number={1},
      pages={16},
      year={2025},
      publisher={Springer}
    }
  3. · Conference paper · Australasian Joint Conference on Artificial Intelligence

    Automated Fish Classification Using Unprocessed Fatty Acid Chromatographic Data: A Machine Learning Approach

    Jesse Wood, Bach Hoai Nguyen, Bing Xue, Mengjie Zhang, Daniel Killeen

    Abstract

    Fish is approximately 40% edible fillet. The remaining 60% can be processed into low-value fertilizer or high-value pharmaceutical-grade omega-3 concentrates. High-value manufacturing options depend on the composition of the biomass, which varies with fish species, fish tissue and seasonally throughout the year. Fatty acid composition, measured by Gas Chromatography, is an important measure of marine biomass quality. This technique is accurate and precise, but processing and interpreting the results is time-consuming and requires domain-specific expertise. The paper investigates different classification and feature selection algorithms for their ability to automate the processing of Gas Chromatography data. Experiments found that SVM could classify compositionally diverse marine biomass based on raw chromatographic fatty acid data. The SVM model is interpretable through visualization which can highlight important features for classification. Experiments demonstrated that applying feature selection significantly reduced dimensionality and improved classification performance on high-dimensional low sample-size datasets. According to the reduction rate, feature selection could accelerate the classification system up to four times.

    BibTeX
    @inproceedings{wood2022automated,
      title={Automated fish classification using unprocessed fatty acid chromatographic data: a machine learning approach},
      author={Wood, Jesse and Nguyen, Bach Hoai and Xue, Bing and Zhang, Mengjie and Killeen, Daniel},
      booktitle={Australasian Joint Conference on Artificial Intelligence},
      pages={516--529},
      year={2022},
      organization={Springer}
    }

Projects