How to tune neural networks for success: AI web training series (part II)

This is the second installment of our special three-part series highlighting Rakuten Symphony Cloud Business Unit President Partha Seetala's “A Comprehensive and Intuitive Introduction to Deep Learning” (CIDL) web training series, which provides an accessible introduction to neural networks for anyone working in the field of AI.

Our last issue of Zero-Touch Telecom focused on Partha’s introduction of the data structures and algorithms that power neural networks. Our second installment explores episode two of Partha’s web training series, which focuses on how neural networks can be optimized and fine-tuned for better accuracy and performance. It covers concepts like input normalization and output functions, revealing how these techniques can be applied by AI practitioners and anyone responsible for implementing AI-driven strategies within complex systems, such as telecom networks.

“If you have tried to understand neural networks before and got overwhelmed, I strongly recommend watching these sessions from Partha,” says Rakuten Symphony CMO Geoff Hollingworth. “They explain the required understanding with subtle differences that are different from what you see elsewhere and enhance the ability for people like me to understand.”

The second episode builds on the concepts covered in episode one by explaining how to further refine AI models to make them more efficient and applicable to real world scenarios.

Let’s dive into the top-level takeaways.

The importance of input normalization

Partha highlights why it is important to ensure input data features are normalized to prevent discrepancies in ranges from affecting the model’s ability to learn efficiently.

He gives the example of determining a home’s value based on square feet, number of bedrooms and lot size.

When data points are scattered over a wide range (like house prices represented in millions of dollars or house sizes expressed in thousands of square feet), the neural network struggles to efficiently process these large differences. Normalization becomes necessary to ensure the network can process the data without bias toward features with larger values.

Centering and standardizing input data using techniques like Z-score normalization ensures all features are on a consistent scale so models can learn more effectively.

Output layer tuning considerations

Partha also explains the role of output functions and loss functions in neural network training and accurate predictions. The output layer is where the network gives its final result. The values coming out of the neurons in this layer need to be transformed into a form that makes sense for a specific task, like predicting a number or classifying a category.

“Tuning” the layer involves selecting the right output function and loss function depending on the task (e.g., regression, classification, ranking, etc.). Various output function types like linear functions for regression tasks, softmax for classification and loss functions like mean squared error and cross-entropy loss are covered, along with examples of when to use each. Once again, Partha referred to the example of predicting house prices to provide a tangible example of when to rely on which method.

Telecom considerations for neural network tuning

For telecom practitioners, these techniques can transform how AI models are developed to accurately and efficiently support network management.

In a telco environment, input normalization is essential due to the wide variation in data from parameters like signal strength, latency and traffic volume. By normalizing and standardizing this data, neural networks can avoid biases caused by differences in scale, leading to more balanced and accurate predictions for tasks such as congestion detection and resource allocation.

On the output side, neural networks can be tuned to deliver practical, actionable insights for telecom-specific use cases like predicting network performance metrics or classifying different types of network anomalies. Selecting the appropriate output function ensures that predictions are tailored to real-world telecom applications.

Fine-tuning both input and output layers can help optimize network performance, enhance customer experiences and drive automation across services.

‍