T9ner/autoresearch-waste

Autonomous AI research for waste classification in Nigeria # Autoresearch for Waste Classification Autoresearch experiment for training and evaluating a computer-vision model that classifies waste images and estimates recoverable material yield. The project is adapted from the autoresearch pattern for waste/recycling use cases, with a focus on practical waste streams relevant to Nigeria and similar contexts. ## What It Does The model looks at an image of waste and predicts one of three categories: - **E-waste** (`class 0`): batteries, phones, electronics, laptops, monitors, computers, TVs - **Recyclable / dry waste** (`class 1`): plastic, bottles, wrappers, glass, metal, paper, cardboard, cans - **Organic waste** (`class 2`): food waste, compostable or biodegradable waste The training script also predicts a rough **yield percentage**: an estimate of how much useful/recoverable material may be available from that waste category. ## Latest Results The latest completed cloud GPU loop was run on May 22 and its metrics are committed in `results.tsv`. Summary from **100 completed validation runs**: | Metric | Min | Average | Max | |---|---:|---:|---:| | `accuracy` | `99.30` | `99.846` | `100.00` | | `yield_mse` | `0.0091` | `0.01020` | `0.0115` | | `combined_score` | `99.30` | `99.846` | `100.00` | | `memory_gb` | `2.87` | `2.87` | `2.87` | Additional notes: - All **100 / 100** experiments completed successfully. - **27 runs** reached `100.00` validation accuracy. - The best observed `yield_mse` was `0.0091`. - The runs used the same committed training setup, so these results mainly show **stability/repeatability** of the current model pipeline rather than a comparison between many different model architectures. - The model used a pretrained computer-vision backbone and was evaluated on the validation split produced by `train.py`. ## Brief Layman Summary In simple terms, we taught a computer to look at pictures of waste and sort them into useful categories: electronics, recyclable materials, or organic waste. We then …