feat: add segment 3 dataset images submission for cataluna84

[cataluna84]

Summary

This PR submits the solution for Segment 3: Dataset Images, focusing on interpreting the mixed4a layer of InceptionV1. It implements a pipeline to correlate dataset examples with feature visualizations to understand neuron behavior.

Technical Details

Notebook: segments/segment_3_dataset_images/submissions/cataluna84__segment_3_dataset_images.ipynb

ML Pipeline & Workflow

1. Model: Pre-trained InceptionV1 (GoogLeNet) via lucent.modelzoo. Target Layer: mixed4a.
2. Dataset Streaming: Efficiently streams ImageNet-1k from HuggingFace, processing ~1.2M images without full disk download.
3. Activation Tracking:
   • Approach: Scans the dataset to find images that maximally (and minimally) activate specific neurons.
   • Spectrum Tracker: Uses ActivationSpectrumTrackerV2 to track 4 categories per neuron: Max Activating, Slightly Positive, Slightly Negative, and Min Activating (Inhibitory).
   • Optimization: Memory-efficient min-heap implementation (SampleRecord) to track top-k samples on the fly.
4. Feature Visualization (Optimization):
   • Uses Lucent (lucent.optvis) to perform gradient ascent on the input space.
   • Generates Positive (Dream) and Negative (Avoidance) synthetic images to compare with real data.
5. Visualization:
   • Outputs a Distill.pub-style Activation Spectrum, arranging synthetic and real images side-by-side to reveal the "ideal" feature vs. real-world triggers.
6. Experiment Monitoring: Integrated with Weights & Biases to log throughput (img/sec) and progress.

Verification

• Confirmed streaming pipeline throughput.
• Verified interaction between ActivationExtractor hooks and the main loop.
• Validated generation of both optimized images and dataset activation tracking.

[cataluna84]

It took 31 hours 33 minutes to go through the whole ImageNet dataset with a batch size of 128 and sample_record_limit as 2000 taking about 20 GB of RAM.

https://wandb.ai/cataluna84/vision-interpretability/runs/ybjw2oe7?nw=nwusercataluna84

Let me know if you are getting the same results with the hyperparameters I am using, and any other questions that you may have.