Creator prompt
The idea behind this presentation
CIVIC -> Create a 14-slide, English-language, widescreen (16:9) academic PowerPoint for BITS F464 Machine Learning at BITS Pilani Dubai Campus. Title: “Machine Learning Waste Classification for Smart Recycling.” Subtitle: “Concept Integration-2 | Model Design and Implementation.” Presenter: Chirudeva Reddy. Course in-charge: Prof. Pranav M. Pawar. Use [Student ID] on the cover so it can be filled in after export.
Follow the slide order below. Slides 1–12 form an approximately eight-minute talk for a machine learning instructor. Slides 13–14 are backup. The LMS requires the problem statement, model architecture and algorithm, results and discussion, and conclusion. Make each visible. Phase 1 framed the problem and proposed the experiments. Phase 2 implemented, tuned and evaluated them.
Design: plain white canvas, charcoal text, blue for RealWaste/local training, warm orange for TrashNet/source training, and muted grey for references and baselines. Use native editable shapes, tables and charts. Use a readable academic sans-serif, large chart labels and one central visual per slide. Keep visible text short and put technical detail into 2–4 sentences of speaker notes per main slide. Use the supplied measurements verbatim. Report averages over five seeds unless a plot is labelled seed 0. Label standard deviations as SD where given. Use only genuine experiment graphics or accurate redraws. The figure filenames below are optional references. If a figure is unavailable, use the supplied values for an editable chart, or leave a labelled placeholder for a real photo, PCA plot, or screenshot. Use numbered references [1]–[7] on relevant slides. Keep the filenames and production instructions in the notes, outside the visible presentation.
1. TITLE — Machine Learning Waste Classification for Smart Recycling
Use the cover information above and one small optional real TrashNet/RealWaste comparison from fig_samples.png. Notes: Ask whether a sorter trained on objects photographed against white posterboard will still work on landfill images. Disclose the evaluation scope as one labelled item per image.
2. PROBLEM STATEMENT — Studio training does not transfer reliably to landfill images
Use an aligned comparison of TrashNet (2,527 studio images, six classes, white posterboard) [1] and RealWaste (4,752 landfill-audit images, nine classes) [2]. Shared materials: cardboard, glass, metal, paper, plastic. Introduce O1 Diagnose the domain gap, O2 Measure how many local labels repair it, and O3 Test when the model can sort automatically. Notes: A separate six-class sensitivity analysis pairs TrashNet “trash” with RealWaste “Miscellaneous Trash.” The five shared material classes are the primary transfer task. Single et al. [3] established a clean-to-real transfer gap. This project connects controlled stress tests, real-label budgeting and stream-level sorting decisions.
3. MODEL ARCHITECTURE — Two input domains, one feature pipeline
Draw a clear branching diagram with editable boxes. TrashNet and RealWaste → stratified splits → parallel handcrafted and frozen-CNN branches → cached feature vectors → StandardScaler → PCA retaining 95% of training variance → LR / RBF-SVM / RF / k-NN. Show the side branch from TrashNet object masks to the O1 stress transforms and the separate RealWaste calibration path to O3. Keep test paths visually distinct from model training. Reference fig_architecture.png. Handcrafted branch: resize to 128 × 128, extract 96 HSV colour, 28 LBP texture and 1,764 HOG shape features. CNN branch: resize to 256, centre-crop to 224 × 224, apply ImageNet normalization and extract a 1,280-dimensional frozen MobileNetV2 embedding [6]. Fusion is 3,168 dimensions. Notes: All 7,279 image files decoded, so no missing-value imputation was required. The frozen backbone kept the experiments feasible on a laptop and supported a direct comparison with handcrafted cues.
4. TRAINING ALGORITHM — Cross-validation chooses the model and calibration sets the routing threshold
Show numbered steps: (1) Stratify 70% train / 15% calibration / 15% test for each of five seeds, 0–4. RealWaste counts are 3,326 / 713 / 713 on the nine-class task. (2) Compare three feature sets (handcrafted, CNN only, fused) and four classifiers under stratified five-fold grid search on the training split, scored by macro-F1. (3) Refit the selected pipeline on all training rows. (4) Evaluate the held-out test split. (5) In O3, compute the conformal threshold on the separate calibration split before scoring the routing rule on test images. Notes: Refit StandardScaler and PCA within each CV fold. TrashNet-only models are tuned on TrashNet folds. Macro-F1 gives equal weight to each class. Report accuracy, macro-precision and macro-recall alongside it. LR, SVM, RF and k-NN supply linear, nonlinear kernel, ensemble and distance-based comparisons.
5. MODEL SELECTION AND BIAS–VARIANCE — The RBF SVM wins the nine-class comparison
Chart fused nine-class RealWaste test macro-F1, mean ± SD: LR 0.845 ± 0.019, RBF-SVM 0.883 ± 0.015, RF 0.712 ± 0.013, k-NN 0.774 ± 0.020. Include a small seed-0 bias–variance panel, or use fig_bias_variance.png: at SVM C = 0.01 train and CV macro-F1 are both about 0.020; at C = 1 CV is 0.852; at C = 10 train is 1.000 and CV is 0.870. A learning curve reaches CV 0.632 at 266 training examples and 0.870 at 2,660. Notes: Small C means stronger regularization and underfitting here. The train–CV gap at C = 10 suggests variance, while the flat validation curve for larger C and the rising learning curve support testing with more data. C = 10 and gamma = scale were chosen on all five nine-class RealWaste seeds. The CNN-only SVM was already at 0.880 ± 0.015, so fusion gave a small gain in that setting.
6. BASELINE RESULTS — Macro-F1 falls from 0.896 to 0.520 on the same RealWaste test images
Use paired bars for the fused RBF SVM over the five shared classes, y-axis “Macro-F1 on RealWaste test.” RealWaste-trained 0.896 ± 0.012; TrashNet-trained 0.520 ± 0.033. Add a small metrics table with columns accuracy / macro-precision / macro-recall / macro-F1: RealWaste-trained 0.895 / 0.910 / 0.887 / 0.896; TrashNet-trained 0.563 / 0.559 / 0.534 / 0.520. Notes: The gap is 0.376 macro-F1. The TrashNet-trained model reached 0.895 in internal TrashNet CV, which differs from held-out RealWaste transfer. Glass was often confused with plastic. Single et al. [3] also reported a large accuracy drop (89.19% real-trained vs 49.69% clean-trained); their split and label protocol differ, so compare the pattern only.
7. O1 METHOD — Change one realism factor while keeping the trained model fixed
Show the original studio image and the five applied factors from fig_stress_gallery.png: background swap, stains, elastic deformation, lighting and occlusion. Indicate severities 1–3. Notes: Fit on clean TrashNet training images. Apply each transform one at a time to held-out TrashNet images and score the same model. For the swap, use Lab colour distance, Otsu threshold and morphology to mask the item, with GrabCut fallback. Use floor patches from RealWaste training images only. Compare stressed versus real off-diagonal confusion patterns and use joint feature-group permutation to inspect cue reliance.
8. O1 FINDINGS — Background swap is the strongest tested factor
Use the fused SVM at severity 3. Chart the stressed TrashNet test means: background 0.634, occlusion 0.752, deformation 0.794, lighting 0.808, stains 0.869. Mark clean TrashNet test 0.877 and actual RealWaste test 0.520 as separately labelled reference lines. Cite fig_o1_drop.png if available. Callout: 0.243 macro-F1 drop, about 68% of the clean-to-real stress gap, using the ratio of rounded mean scores. Notes: The gap-share denominator is clean TrashNet test minus RealWaste test for the same TrashNet-trained model. Factors were tested in isolation, their shares can overlap, and none reached 0.520 individually. Background ranked first across all four classifiers with fused features. For the fused SVM, the background-swap error pattern correlates with RealWaste at r = 0.57. The use of RealWaste floor pixels and imperfect masks limit causal attribution. Qin et al. [4] provide related evidence on background and saliency.
9. O2 FINDINGS — Approximately 23 local labels per class recover half the gap
Draw a line chart using the following mean macro-F1 values on the five-class RealWaste test. X-axis: real training labels per class k = 0, 5, 10, 25, 50. TrashNet + real, no augmentation: 0.520, 0.611, 0.635, 0.717, 0.778. Generic augmentation: 0.500, 0.605, 0.629, 0.714, 0.769. Targeted augmentation: 0.552, 0.596, 0.620, 0.689, 0.755. Real-only: missing at k = 0, then 0.428, 0.542, 0.697, 0.776. Dashed full real-training reference 0.896. Highlight 50 labels/class = 250 images, 69% of the gap recovered. Notes: Local samples were nested within each seed, and each arm was tuned again. Gap recovery = (adapted minus 0.520) / (0.896 minus 0.520). The half-gap budget is interpolated, and no arm reached 80% recovery by k = 50. Targeted augmentation gave +0.032 ± 0.030 at k = 0 (paired, 4 of 5 seeds, p ≈ 0.08) and trailed no augmentation at k ≥ 5. Generic augmentation gave no gain at the tested budgets. Compare cautiously with Babé et al. [5].
10. O3 ALGORITHM — Sort only when every plausible class belongs to one stream
Draw two linked paths. Calibration path: train a nine-class fused SVM with probability outputs on RealWaste train → use 713 untouched calibration images to compute s = 1 − p(true class) → choose the finite-sample conformal threshold q at level α. Inference path: collect classes satisfying 1 − p(class) ≤ q → map to dry recyclables, organics, or residual → auto-sort if the set is nonempty and contains exactly one stream; otherwise defer to a human sorter. Include two labelled schematic examples: {paper, cardboard} → dry recyclables; {plastic, food organics} → manual sorting. Notes: The quantile uses the ceiling rank ceil((n + 1)(1 − α)), with conservative rounding as implemented. Under exchangeability the true label belongs to the class set with marginal probability at least 1 − α [7]. Alpha governs class-set coverage. Wrong-stream rate is measured empirically among automatically sorted items.
11. O3 RESULTS — 90.7% automation with a 1.50% wrong-stream rate
Use a labelled scatterplot or compact table. Always-sort top-1: automation 100%, wrong-stream 3.56%. α = 0.10: 97.3%, 2.68%. α = 0.05: 90.7%, 1.50%. α = 0.02: 79.4%, 0.60%. α = 0.01: 66.5%, 0.49%. Label the x-axis “Items automatically sorted (%)” and y-axis “Wrong stream among sorted items (%)” if charting. Highlight α = 0.05: 95.1% observed class-set coverage and 9.3% deferred. Notes: These are held-out, in-domain RealWaste averages across five seeds. At this point, automation is 90.69% ± 2.52 percentage points SD and wrong-stream rate is 1.50% ± 0.85 percentage points SD. The top-1 comparator automatically routes every item and already meets the Phase 1 5% in-domain target. Conformal sorting helps when stricter misrouting is worth the additional manual workload. The guarantee is marginal and depends on exchangeability, and new sites need their own calibration and testing.
12. CONCLUSION — Control the background, collect local labels, then calibrate locally
Use three restrained takeaway boxes, one each for O1, O2 and O3. Include the headline numbers 0.243 macro-F1 drop, about 69% gap recovery with 50 labels/class, and 90.7% automation at 1.50% wrong-stream. Add a short “Limitations and next steps” line: stress combinations, more labels and sites, class-conditional calibration, and a dashboard user study. Notes: The results are from image datasets, and a field deployment has yet to be evaluated. The experiment does not estimate the gain from physically changing a sorting line. Miscellaneous Trash is a weak class. Finish by explaining how a new site could use these findings to prioritize what to test.
13. BACKUP — Implementation, dashboard and corroborating analysis
Show the three Streamlit tabs: Diagnose an image, Sort or defer, Dataset results. Use an actual screenshot if supplied, otherwise a simple labelled diagram. Notes: The app supports original vs stressed input, prediction probabilities, occlusion sensitivity, feature-group ablation and an α slider. Run with “streamlit run app.py.” Code uses Python, scikit-learn, PyTorch/torchvision, OpenCV, scikit-image, NumPy, pandas, Matplotlib and Streamlit. Results, dataset download instructions, dependencies and run steps are documented in the repository. The dashboard was built; a user study remains pending. Fused k-means adjusted Rand index against dataset vs material: 0.717 vs 0.018 at k = 2, and 0.404 vs 0.127 at k = 5. The k = 2 comparison favours the binary domain label. Clustering measures representation structure and does not establish transfer performance.
14. BACKUP — Selected references
[1] G. Thung and M. Yang, TrashNet dataset, 2016. github.com/garythung/trashnet
[2] S. Single, S. Iranmanesh, and R. Raad, RealWaste dataset, UCI Machine Learning Repository, 2023. doi:10.24432/C5SS4G
[3] S. Single, S. Iranmanesh, and R. Raad, “RealWaste: A Novel Real-Life Data Set for Landfill Waste Classification Using Deep Learning,” Information, 2023. doi:10.3390/info14120633
[4] J. Qin et al., “A robust framework combined saliency detection and image recognition for garbage classification,” Waste Management, 2022. doi:10.1016/j.wasman.2021.11.027
[5] A. Babé et al., “Generalization abilities of foundation models in waste classification,” Waste Management, 2025. doi:10.1016/j.wasman.2025.02.032
[6] M. Sandler et al., “MobileNetV2: Inverted Residuals and Linear Bottlenecks,” CVPR, 2018. doi:10.1109/CVPR.2018.00474
[7] A. N. Angelopoulos and S. Bates, “Conformal Prediction: A Gentle Introduction,” Foundations and Trends in Machine Learning, 2023. doi:10.1561/2200000101
Before finishing, check that the technical story keeps the five-class transfer and label-budget experiments separate from the nine-class, three-stream experiment. Keep class coverage, accuracy, macro-F1, and wrong-stream rate distinct. Keep the calibration split out of model tuning. Preserve the weak augmentation result and the same-test-set comparison. Export to PowerPoint.