Scribd
Upload
10 views7 pages

ANN for Digit Recognition with SVHN Dataset

The document outlines a project on implementing an Artificial Neural Network (ANN) for digit recognition using the SVHN dataset, detailing data preparation, model architectures, and training processes. Two models were developed, with the optimized ANN achieving a 77% accuracy, which is better than Logistic Regression but less effective than CNNs. The conclusion highlights the limitations of ANN for image data and suggests future work involving CNNs to improve accuracy by retaining spatial features.

Uploaded by

sachinmishra202002
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
10 views7 pages

ANN for Digit Recognition with SVHN Dataset

The document outlines a project on implementing an Artificial Neural Network (ANN) for digit recognition using the SVHN dataset, detailing data preparation, model architectures, and training processes. Two models were developed, with the optimized ANN achieving a 77% accuracy, which is better than Logistic Regression but less effective than CNNs. The conclusion highlights the limitations of ANN for image data and suggests future work involving CNNs to improve accuracy by retaining spatial features.

Uploaded by

sachinmishra202002
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd

Artificial Neural Network (ANN) -

Project Implementation
• Topic: ANN for Digit Recognition using SVHN Dataset

• Code Focus: Model Definition, Training, and Evaluation


• By: [Your Name]
• Course: Master of Computer Application
Data Preparation & Setup
• • Keras backend cleared to ensure reproducible results.
• • Random seeds fixed (NumPy, TensorFlow, random) for consistency.
• • Dataset reshaped into flattened form (32×32 = 1024 input features).
• • One-hot encoding applied to digit labels (0–9).
• • Final input shape: (samples, 1024) | Output shape: (samples, 10).
Model 1: Simple ANN
• • Architecture:
• - Input Layer: 1024 neurons
• - Hidden Layer 1: Dense(64), activation='relu'
• - Hidden Layer 2: Dense(32), activation='relu'
• - Output Layer: Dense(10), activation='softmax'

• • Optimizer: Adam (learning rate=0.001)


• • Loss: Categorical Crossentropy | Metric: Accuracy
• • Epochs: 20 | Batch Size: 128
Model 2: Optimized ANN
(Improved Architecture)
• • Architecture:
• - Input: 1024 neurons
• - Dense(256) → Dense(128) → Dropout(0.2)
• - Dense(64) → Dense(64) → Dense(32)
• - BatchNormalization() → Output(10, Softmax)

• • Optimizer: Adam (learning rate=0.0005)


• • Loss: Categorical Crossentropy | Metric: Accuracy
• • Epochs: 30 | Batch Size: 128
Training & Accuracy Tracking
• • Accuracy plotted for both training and validation sets.
• • Accuracy curve stabilizes around 77% for Model 2.
• • Validation accuracy consistent with training accuracy, indicating no overfitting.
• • Graph shows steady improvement across 30 epochs.
Model Evaluation & Results
• • Model 2 predictions compared with actual test labels.
• • Classification Report: 77% overall test accuracy.
• • Confusion Matrix visualized using seaborn heatmap.
• • Model performs better than Logistic Regression (13%) but weaker than CNN
(90%+).
Final Observations & Conclusion
• • ANN achieved 77% accuracy, proving effective but limited for image-based data.
• • Balanced training and validation accuracy shows good generalization.
• • Common misclassifications: digits 3, 5, 6, and 8.
• • ANN loses spatial context due to flattened input.
• • Future Work: Use CNN to retain spatial features for higher accuracy.

You might also like