Figure Gallery
66 figures from the thesis, filterable by chapter.
Ch. 2: Literature Survey
Mass Spectrograph: This is the artifact generated by Rapid Evaporative Ionization Mass Spectrometry (REIMS). The x-axis
Mass Spectrograph: This is the artifact generated by Rapid Evaporative Ionization Mass Spectrometry (REIMS). The x-axis
Ch. 2: Literature Survey
Illustrated here is a comparison of Transformer models, with each circle's size representing its approximate number of p
Illustrated here is a comparison of Transformer models, with each circle's size representing its approximate number of p
Ch. 2: Literature Survey
This figure illustrates the Scaled Dot-Product Attention mechanism , a core component of self-attention. It takes querie
This figure illustrates the Scaled Dot-Product Attention mechanism , a core component of self-attention. It takes querie
Ch. 2: Literature Survey
Multi-head attention enhances a model's ability to capture complex relationships by allowing it to jointly focus on dis
Multi-head attention enhances a model's ability to capture complex relationships by allowing it to jointly focus on dis
Ch. 3: Datasets and Processing
Mass Spectrograph: This is the artifact generated by Rapid Evaporative Ionization Mass Spectrometry (REIMS). The x-axis
Mass Spectrograph: This is the artifact generated by Rapid Evaporative Ionization Mass Spectrometry (REIMS). The x-axis
Ch. 3: Datasets and Processing
Mackerel — Mackerel (left), Hoki (right) fish species.
Mackerel — Mackerel (left), Hoki (right) fish species.
Ch. 3: Datasets and Processing
Hoki — Mackerel (left), Hoki (right) fish species.
Hoki — Mackerel (left), Hoki (right) fish species.
Ch. 3: Datasets and Processing
This diagram illustrates a machine learning model running on a computer to identify fish species. First, the fish is sam
This diagram illustrates a machine learning model running on a computer to identify fish species. First, the fish is sam
Ch. 3: Datasets and Processing
Fish body parts.
Fish body parts.
Ch. 3: Datasets and Processing
This diagram illustrates a machine learning model running on a computer to identify fish body parts. Similar to the prev
This diagram illustrates a machine learning model running on a computer to identify fish body parts. Similar to the prev
Ch. 3: Datasets and Processing
Fish are exposed to oil contamination from both human and natural sources. Major human-caused sources include oil spills
Fish are exposed to oil contamination from both human and natural sources. Major human-caused sources include oil spills
Ch. 3: Datasets and Processing
Fish can be affected by cross-species adulteration, where lower-value or unrelated species are intentionally or accident
Fish can be affected by cross-species adulteration, where lower-value or unrelated species are intentionally or accident
Ch. 3: Datasets and Processing
This diagram illustrates the task of batch detection for marine biomass. For simplification, we give an example with onl
This diagram illustrates the task of batch detection for marine biomass. For simplification, we give an example with onl
Ch. 4: Fish Species and Part Identifi
This is the transformer architecture proposed in this thesis. It is an encoder-only architecture , with four encoder lay
This is the transformer architecture proposed in this thesis. It is an encoder-only architecture , with four encoder lay
Ch. 4: Fish Species and Part Identifi
This is the MoE Transformer architecture proposed in this thesis. It consists of an encoder-only architecture , with fo
This is the MoE Transformer architecture proposed in this thesis. It consists of an encoder-only architecture , with fo
Ch. 4: Fish Species and Part Identifi
This figure shows the architecture for the stacked voting ensemble classifier, simply referred to here as the Ensemble T
This figure shows the architecture for the stacked voting ensemble classifier, simply referred to here as the Ensemble T
Ch. 4: Fish Species and Part Identifi
LIME explanation for pre-trained transformer for classification of fish species Mackerel.
LIME explanation for pre-trained transformer for classification of fish species Mackerel.
Ch. 4: Fish Species and Part Identifi
LIME explanation for pre-trained transformer for classification of fish species Hoki.
LIME explanation for pre-trained transformer for classification of fish species Hoki.
Ch. 4: Fish Species and Part Identifi
Decision tree for fish species.
Decision tree for fish species.
Ch. 4: Fish Species and Part Identifi
Grad-CAM for pre-trained transformer for classification of fish species.
Grad-CAM for pre-trained transformer for classification of fish species.
Ch. 4: Fish Species and Part Identifi
LIME explanation for transformer for classification of fish part head.
LIME explanation for transformer for classification of fish part head.
Ch. 4: Fish Species and Part Identifi
LIME explanation for transformer for classification of fish part fillet.
LIME explanation for transformer for classification of fish part fillet.
Ch. 4: Fish Species and Part Identifi
LIME explanation for transformer for classification of fish part liver.
LIME explanation for transformer for classification of fish part liver.
Ch. 4: Fish Species and Part Identifi
LIME explanation for transformer for classification of fish part skins.
LIME explanation for transformer for classification of fish part skins.
Ch. 4: Fish Species and Part Identifi
LIME explanation for transformer for classification of fish part guts.
LIME explanation for transformer for classification of fish part guts.
Ch. 4: Fish Species and Part Identifi
LIME explanation for transformer for classification of fish part frames.
LIME explanation for transformer for classification of fish part frames.
Ch. 4: Fish Species and Part Identifi
LIME explanation for transformer for classification of fish part gonads.
LIME explanation for transformer for classification of fish part gonads.
Ch. 4: Fish Species and Part Identifi
Grad-CAM for transformer for classification of fish body part.
Grad-CAM for transformer for classification of fish body part.
Ch. 4: Fish Species and Part Identifi
In this diagram, we present the top 10 features for fish species classification for each method that have been identifie
In this diagram, we present the top 10 features for fish species classification for each method that have been identifie
Ch. 4: Fish Species and Part Identifi
In this diagram, we present the top 10 features for fish part classification for each method that have been identified b
In this diagram, we present the top 10 features for fish part classification for each method that have been identified b
Ch. 4: Fish Species and Part Identifi
This stacked bar chart shows the expert utilization of the MoE Transformer with 4 experts across all four classification
This stacked bar chart shows the expert utilization of the MoE Transformer with 4 experts across all four classification
Ch. 4: Fish Species and Part Identifi
This radar chart illustrates the utilization patterns of four different experts across four task categories: Species, Pa
This radar chart illustrates the utilization patterns of four different experts across four task categories: Species, Pa
Ch. 4: Fish Species and Part Identifi
Majority Voting versus Top-k Routing Bar Chart. This chart gives the test classification accuracy on the fish species an
Majority Voting versus Top-k Routing Bar Chart. This chart gives the test classification accuracy on the fish species an
Ch. 4: Fish Species and Part Identifi
Expert Count Analysis Bar Chart. This bar chart illustrates the test classification accuracy across the two classificati
Expert Count Analysis Bar Chart. This bar chart illustrates the test classification accuracy across the two classificati
Ch. 4: Fish Species and Part Identifi
Top-k Routing Bar Chart. This bar chart illustrates the test classification accuracy for the two tasks of fish species c
Top-k Routing Bar Chart. This bar chart illustrates the test classification accuracy for the two tasks of fish species c
Ch. 5: Oil Contamination and Cross-Sp
Masked Spectra Modelling is a variation of Masked Language Modeling from BERT . But unlike BERT, which uses a bidirectio
Masked Spectra Modelling is a variation of Masked Language Modeling from BERT . But unlike BERT, which uses a bidirectio
Ch. 5: Oil Contamination and Cross-Sp
Transfer Learning Overview: The image outlines the transfer learning workflow with MoE Transformers. Step 1 involves pre
Transfer Learning Overview: The image outlines the transfer learning workflow with MoE Transformers. Step 1 involves pre
Ch. 5: Oil Contamination and Cross-Sp
The confusion matrix on the test dataset for one of the folds of cross-validation. The diagonal of the matrix denotes co
The confusion matrix on the test dataset for one of the folds of cross-validation. The diagonal of the matrix denotes co
Ch. 5: Oil Contamination and Cross-Sp
The prediction error histogram denotes how far off each of the predictions was. Since this is an ordinal classification
The prediction error histogram denotes how far off each of the predictions was. Since this is an ordinal classification
Ch. 5: Oil Contamination and Cross-Sp
LIME explanation for pre-trained transformer for classification of oil contamination in 50\% concentration.
LIME explanation for pre-trained transformer for classification of oil contamination in 50\% concentration.
Ch. 5: Oil Contamination and Cross-Sp
LIME explanation for pre-trained transformer for classification of oil contamination in 25\% concentration.
LIME explanation for pre-trained transformer for classification of oil contamination in 25\% concentration.
Ch. 5: Oil Contamination and Cross-Sp
LIME explanation for pre-trained transformer for classification of oil contamination in 10\% concentration.
LIME explanation for pre-trained transformer for classification of oil contamination in 10\% concentration.
Ch. 5: Oil Contamination and Cross-Sp
LIME explanation for pre-trained transformer for classification of oil contamination in 5\% concentration.
LIME explanation for pre-trained transformer for classification of oil contamination in 5\% concentration.
Ch. 5: Oil Contamination and Cross-Sp
LIME explanation for pre-trained transformer for classification of oil contamination in 1\% concentration.
LIME explanation for pre-trained transformer for classification of oil contamination in 1\% concentration.
Ch. 5: Oil Contamination and Cross-Sp
LIME explanation for pre-trained transformer for classification of oil contamination in 0.1\% concentration.
LIME explanation for pre-trained transformer for classification of oil contamination in 0.1\% concentration.
Ch. 5: Oil Contamination and Cross-Sp
LIME explanation for pre-trained transformer for classification of oil contamination in 0\% concentration.
LIME explanation for pre-trained transformer for classification of oil contamination in 0\% concentration.
Ch. 5: Oil Contamination and Cross-Sp
LIME explanation for pre-trained transformer for cross-species contamination of Hoki-Mackerel class.
LIME explanation for pre-trained transformer for cross-species contamination of Hoki-Mackerel class.
Ch. 5: Oil Contamination and Cross-Sp
LIME explanation for pre-trained transformer for cross-species contamination of Hoki class.
LIME explanation for pre-trained transformer for cross-species contamination of Hoki class.
Ch. 5: Oil Contamination and Cross-Sp
LIME explanation for pre-trained transformer for cross-species contamination of Mackerel class.
LIME explanation for pre-trained transformer for cross-species contamination of Mackerel class.
Ch. 5: Oil Contamination and Cross-Sp
In this diagram, we present the top 10 features for oil contamination classification that have been identified by our 1D
In this diagram, we present the top 10 features for oil contamination classification that have been identified by our 1D
Ch. 5: Oil Contamination and Cross-Sp
In this diagram, we present the top 10 features for fish cross-species adulteration classification that have been identi
In this diagram, we present the top 10 features for fish cross-species adulteration classification that have been identi
Ch. 5: Oil Contamination and Cross-Sp
This stacked bar chart shows the expert utilization of the MoE Transformer with 4 experts across all four classification
This stacked bar chart shows the expert utilization of the MoE Transformer with 4 experts across all four classification
Ch. 5: Oil Contamination and Cross-Sp
This radar chart illustrates the utilization patterns of four different experts across four task categories: Species, Pa
This radar chart illustrates the utilization patterns of four different experts across four task categories: Species, Pa
Ch. 5: Oil Contamination and Cross-Sp
Majority Voting versus Top-k Routing Bar Chart
Majority Voting versus Top-k Routing Bar Chart
Ch. 5: Oil Contamination and Cross-Sp
Expert Count Analysis Bar Chart
Expert Count Analysis Bar Chart
Ch. 5: Oil Contamination and Cross-Sp
Top-k Routing Bar Chart
Top-k Routing Bar Chart
Ch. 5: Oil Contamination and Cross-Sp
Test Classification Improvements Radar Chart: This radar chart provides a comparative view of Baseline (green) and Trans
Test Classification Improvements Radar Chart: This radar chart provides a comparative view of Baseline (green) and Trans
Ch. 5: Oil Contamination and Cross-Sp
A conceptual comparison of data measurement scales, illustrating the unique nature of (b) Ordinal Data, where categories
A conceptual comparison of data measurement scales, illustrating the unique nature of (b) Ordinal Data, where categories
Ch. 5: Oil Contamination and Cross-Sp
Conceptual difference in evaluation. (a) Standard accuracy treats all misclassifications as equally incorrect (a 0-1 los
Conceptual difference in evaluation. (a) Standard accuracy treats all misclassifications as equally incorrect (a 0-1 los
Ch. 5: Oil Contamination and Cross-Sp
Common deep learning strategies for ordinal classification, showing the two approaches tested: (a) The Ordinal Output La
Common deep learning strategies for ordinal classification, showing the two approaches tested: (a) The Ordinal Output La
Ch. 6: Contrastive Learning for Batch
SpectroSim Architecture: Paired samples \(x_1\) and \(x_2\) (REIMS spectra) are processed by identical Transformer
SpectroSim Architecture: Paired samples \(x_1\) and \(x_2\) (REIMS spectra) are processed by identical Transformer
Ch. 6: Contrastive Learning for Batch
This figure illustrates a good ``semantic representation". Imagine that the different colors of fish represent fish from
This figure illustrates a good ``semantic representation". Imagine that the different colors of fish represent fish from
Ch. 6: Contrastive Learning for Batch
Visualization of the six data augmentation techniques described in Section 6.2.2. The Original Spectrum (top left) is sh
Visualization of the six data augmentation techniques described in Section 6.2.2. The Original Spectrum (top left) is sh
Ch. 6: Contrastive Learning for Batch
In the proposed SpectroSim model, we replace the existing ResNet backbone with a Transformer backbone. The choice of tra
In the proposed SpectroSim model, we replace the existing ResNet backbone with a Transformer backbone. The choice of tra
Ch. 6: Contrastive Learning for Batch
This diagram illustrates the difference between binary classification and contrastive learning. Here, we see binary clas
This diagram illustrates the difference between binary classification and contrastive learning. Here, we see binary clas
Ch. 6: Contrastive Learning for Batch
Average Grad-CAM for our Transformer-based SpectroSim model. The visualization highlights the most salient m/z features
Average Grad-CAM for our Transformer-based SpectroSim model. The visualization highlights the most salient m/z features