Deep Learning: Enhancing your Predictive Models with High Level Representations

Problem Statement

Traditional data mining algorithms attempt to search for a predictive model by looking at raw variables that often bear little correlation with the class for prediction.

As an example, predicting stuck-pipes during drilling can be extremely complicated when one looks at raw variables that exhibit high interactions and carry little relevant information on their own.

A similar situation occurs during reservoir characterization when the goal is to automatically identify geological bodies from seismic data. Features extracted from segmenting data cubes are often insufficient to identify entire geological bodies accurately and with few misclassifications.

A new approach is required to address this type of situations.

 

Deep learning

Deep learning is now considered a revolutionary approach to machine learning and data mining. The idea is to enhance predictive models by finding “high-level representations‚ÄĚ. High level representations are new variables automatically constructed by the learning algorithm that try to capture highly informative and meaningful information regarding the object under analysis. As an example, raw pixels in an image can serve as input to a deep learning architecture; the result is a new set of variables that capture edges, contours, and relevant patterns to perform fast and efficient object recognition.

  • Deep learning provides a complete new methodology to enhance the quality of the predictive models by searching for complex combinations of variables describing the problem under analysis.
  • By constructing tenths of layers of abstract new features, or high-level representations, deep learning attempts to disentangle the complex interaction among variables until the problem representation is amenable to data modeling.
  • Deep learning offers a new set of highly informative variables, and more accurate predictive models.

 

Implementation

Noesis provides consulting services to train technical personnel in the use of deep learning techniques for data analysis and model prediction. Training involves using the most advanced software tools implementing deep learning architectures.

We also provide specialized software solutions that address specific requirements of data mining that can be greatly enhanced using deep learning architectures.

Automatic Assistants to Select Data Mining and Machine Learning Techniques

Problem Statement

Data mining and machine learning techniques have become pillars during the analysis of massive amounts of data. Such fields of study have generated hundreds of different algorithms for generation of predictive models, finding patterns in data, automatically classifying objects or events into new categories, etc.

Nevertheless, the successful use of these tools outside the boundaries of research (e.g., industry, commerce, government) is conditioned on the appropriate selection of a suitable predictive model (or combinations of models) according to the domain of application. Without any kind of assistance, model selection and combination can turn into stumbling blocks to the end-user who wishes to access the technology more directly and cost-effectively. End-users often lack not only the expertise necessary to select a suitable model, but also the availability of many models to proceed on a trial-and-error basis.

 

Automatic Assistants

Noesis solution to this problem is to provide automatic assistants and meta-learning systems. These systems can provide automatic and systematic user guidance by mapping a particular task to a suitable model (or combination of models). We analyze properties of the data under analysis, and suggest data analysis strategies that best match the strength of specific data mining tools. This approach greatly expedites the effort in finding the right set of learning algorithms and parameters to successfully implement data mining and machine learning solutions.

 

Implementation

Noesis provides software solutions to automatically select predictive models from a model repository made of hundreds of different approaches.

Our machine learning experts provide consulting services by pointing to the strengths and weaknesses of different methods, closely interacting with the clients, suggesting the best approach to model selection and combination.

Noesis services encompass the whole sequence of steps needed to deploy a predictive model or to find patterns in data. We provide consulting services during data collection, processing, cleaning, data transformation, finding patterns or predictive models, interpretation and evaluation, and generation of new knowledge.

Predictive Analytics

Problem Statement

Our times are characterized by an unprecedented flood of data streams, together with a lack of efficient tools to extract useful pieces of information from massive data repositories. Applications such as credit rating, medical diagnosis, machine failure prediction, fraud detection, and identification of astronomical objects in images, generate thousands of instances for analysis at an unprecedented rate, outpacing the technology needed for proper analysis and interpretation.

 

At the core of modern analysis tool is the generation of “predictive models” where computers automatically sift through huge volumes of data to produce models that classify objects or events into a hierarchy of categories. This is extremely useful for many real-world applications where accurate predictions can easily lead to an increase in productivity and operational efficiency.

 

But despite the availability of many software tools incorporating data analytics software, such packages provide little or no additional information about the type of analysis technique most appropriate for the task at hand. In addition, real-world applications are generally time-sensitive, and in need of specialized data analysis solutions.

 

Predictive Analytics

Noesis methodology consists of a detailed and exhaustive analysis of project goals to decide what data analysis tools best fit the problem at hand. Our experts combine notions in probability and statistics, machine learning, optimization, and advanced mathematical techniques to produce a solution package that addresses each of the client’s needs.

Many of our products are often the combination of several techniques that together provide a comprehensive solution to a very specific data analytics problem. As an example, a particular case study looking for the accurate prediction of stuck pipes during drilling revealed that the best approach is to form an “ensemble” of different techniques (including neural networks and decision trees). Such combinations are hard to discover, often requiring an exhaustive search within a huge space of algorithm combinations.

 

Implementation

Our company provides consulting services to help along the whole data analysis process. We address all steps necessary for the success of a data analytics solution, including the data collection, data cleaning and normalization, searching for patterns using advanced machine learning techniques and high performance computing, and aiding in data interpretation and visualization.

The goal is to help our clients from the very definition of a problem to the delivery of an efficient and cost-effective solution.