Neural Networks

Today, neural networks are used to solve a wide variety of problems, some of which have been solved by existing statistical methods, and some of which have not. Neural networks can provide more accurate and robust solutions for problems where traditional methods fails to provide adequate quality. Neural networks applications fall into one of the following three categories:
  • Forecasting: predicting one or more quantitative outcomes from both quantitative and categorical input data,
  • Classification: classifying input data into one of two or more categories, or
  • Pattern recognition: uncovering patterns, typically spatial or temporal, among a set of variables.
When building neural networks, one of the main goals is to model such aspects of the human nervous system as the ability to learn and fix mistakes. A neural network is able to learn on its own and act based on prior experience, making fewer and fewer mistakes. Neural networks have demonstrated excellent results in image recognition and solutions for other tasks which could not be solved at this level using other mathematical methods.

Bell.One offers our clients neural network development as a solution to different kinds of problems. We have experience developing and implementing machine vision systems using neural networks on both our proprietary architecture and using other productive architectures on the market.

For example, we've created a service capable of recognizing different models/production years/ of car at 96% accuracy, which is in the range of and even exceeds that of an educated person.

Use Cases

Use cases for Neural Networks are derived from four major areas:
  • Approximation
  • Forecasting
  • Classification
  • Clustering


This is a very basic task which as a basic step is encountered almost in any analytical task. It is almost nonexistent in its pure form.
  • Missing data. Restoration of data that was lost due to some technical or administrative problems. Meteorological sensors gathering weather data in remote locations with harsh conditions.
  • Approximate count. Consider 50 millions online reviews from an online retail vendor, with the task to approximate the number of customers behind these reviews. Customers write multiple reviews, of unknown quantity.
  • Approximate quantiles. Insurance company wants to estimate claim experts load. Most claims are expressed in few hundred words.But some are masterpiece literature. The aggregate information such as the mean, the variance, the min, the max, and the percentiles could be approximated without exact computation.


Forecasting is an extension of approximation. The main difference is that in this case, we aim to predict future data in time.
  • Sales forecasting
  • Commodity price prediction
  • Customer life time value prediction
  • Machinery failure prediction. Maintenance on demand instead of time interval.


Neural Networks can be used effectively to classify samples, i.e., map input data to different classes or categories. The renaissance of this technology came when the Neural Networks showed overwhelmingly better results on image classification problems – traditionally very hard to solve. The car recognition service mentioned at the beginning of this article is the classic use case for Neural Networks.
  • Image recognition of any kind. Content based image retrieval systems. Deduplication. Face recognition. Emotion analysis. Customer counting in “brick and mortar” retail. Object identification in security and autopilots. Digital merchandising. Etc.
  • Voice recognition. Services like Shazam and personal assistant like Amazon Echo
  • Text classification from simple spam filters to advanced claim classifier for insurance company. Sentimental analysis. Text summarization
  • Recommendation services based on customers or products similarities
  • Credit risk modelling


Clustering is another form of classification, where the numbers of classes are not known beforehand. Therefore the working of neural networks for clustering is similar to classifying records.
  • Fraud detection
  • Customers micro segmentation for precision marketing
  • Cell tower placement by clustering customers for optimal signal strength
  • Loyalty card customer’s segmentation.

Key Features

  • Have the ability to learn and model non-linear and complex relationships
  • Can generalize — After learning from the initial inputs and their relationships, it can infer unseen relationships on unseen data as well, thus making the model generalize and predict on unseen data
  • Neural Networks can be applied to a variety of problems
  • Neural Networks can be used to solve complex problems that cannot be solved using statistical techniques
  • Unlike many other prediction techniques, Neural Networks does not impose any restrictions on the input variables
  • Readily available frameworks for enterprise level production

How It Works

Although neural network solutions to forecasting, pattern recognition and classification problems can be very different, they are always the result of computations that proceed from the network inputs to the network outputs. The network inputs are referred to as patterns, and outputs are referred to as classes.


  • Much better quality of results in broad variety of application, especially pattern recognition
  • Feature engineering is automated. Cost reduction on high payroll staff.