Learning new tasks without new data.
This project explores how machine learning models themselves can be generated using other machine learning models, without direct task-specific supervision.

This project introduces an unconventional but powerful idea:
using generative models to create new classifiers by fusing existing ones.

Specifically, we leverage CycleGANs to merge the learned feature representations of two independently trained Convolutional Neural Networks (CNNs):

• One CNN trained to recognize cats
• One CNN trained to recognize the color black

By translating and combining their feature spaces, we generate a new CNN capable of detecting black cats — without ever being trained on a single black cat image.

📊 Result:
The generated model achieves 88% classification accuracy on black cat detection, despite zero direct exposure to black cat data.

This work opens new directions for:

• Learning under data scarcity
• Automated model generation
• Knowledge transfer beyond traditional fine-tuning

• 🔁 CycleGAN-based Model Fusion
  Uses CycleGANs to translate and align feature kernels between CNNs trained on unrelated domains.

• 🧬 Generated CNNs (Zero-shot Task Creation)
  Constructs a task-specific classifier purely from pre-trained models.

• 📉 Feature Space Validation
  Employs UMAP, K-Means, and DBSCAN to analyze and validate learned representations.

• 🧪 Unsupervised Generalization
  Demonstrates black cat recognition without labeled black cat data.

Instead of training a model on data, we train a model on other models.

1. Train two CNNs on separate concepts

   • Object: cat
   • Attribute: black

2. Extract convolutional kernels from both networks.

3. Train CycleGANs to translate kernels between these feature domains.

4. Initialize a new CNN using the CycleGAN-generated kernels.

5. Evaluate whether this synthesized CNN can recognize black cats.

➡️ It can.

• 2 Convolutional layers
• Kernel size: 5×5
• Activation: ReLU
• Max-pooling layers
• Trained independently on separate domains

• Generator–Discriminator architecture
• Learns kernel-space translation, not image translation
• Operates directly on convolutional filters

• Initialized entirely using CycleGAN-generated kernels
• No gradient updates using black cat images

• Accuracy
• Precision & Recall
• Cluster Entropy
• Cluster Purity
• Cosine Similarity
• UMAP Visualization

• ✅ The generated CNN successfully clusters black cat images
• 📉 UMAP projections show clear separation of semantic concepts
• 📐 Cosine similarity confirms meaningful feature alignment
• 🧠 Demonstrates unsupervised semantic composition

The model learns “black AND cat” without ever seeing a black cat.

Traditional ML assumes:

New task ⇒ new labeled data

This project challenges that assumption by showing:

• Tasks can be composed
• Models can be generated, not trained
• Generative models can operate in parameter space, not just data space

This has implications for:

• Low-resource domains
• Privacy-sensitive data
• Automated ML systems
• Foundation model composition

• 🔧 Hyperparameter optimization for clustering and feature fusion
• 📐 Alternative feature-space similarity metrics
• 🧠 Semantic-aware end-to-end pipelines
• 🧪 Scaling to deeper CNNs and transformers
• 🔁 Multi-attribute model composition

• GPU with ≥ 6GB VRAM (recommended)

• Python 3.11
• PyTorch 2.5
• torchvision
• numpy
• scikit-learn
• seaborn
• matplotlib
• tqdm

Install dependencies:

pip install -r requirements.txt

###📄 License MIT

Generative models, representation learning, and non-traditional ML paradigms.

“Why train on more data when you can train on more models?”