Deep Learning Algorithms: Deep Learning Through TensorFlow

This article was written by David Berger, a Financial Analyst at I Know First and studying Finance at the University of Michigan’s Ross School of Business.

Deep Learning Algorithms: Deep Learning Through TensorFlow

Deep Learning Algorithms


  • What is Deep Learning?
  • TensorFlow- Google’s Open Source Software Library
  • TensorFlow Compared to Other Libraries
  • Early Results from TensorFlow

What is Deep Learning?

Deep Learning Algorithms

While new innovative products such as Amazon’s Alexa seemingly coming out every day, many consumers wonder how such technology works. For most, they just know the terms machine learning and deep learning. But what do those terms truly mean? Deep learning is the use of multiple artificial neural networks that contain at least one hidden neural layer. These artificial neural networks use non-linear processing units to create information that builds on itself. “Layers” form and the network creates a hierarchy of concepts to better understand new information.  A lone neuron might take in various inputs with assigned weights and output an answer, and other neurons connect to form a neural net. A neural net uses its hierarchy of concepts to assigned inputs and outputs specific weights to produce a higher quality output.

But how are the different input weights assigned? Assignment is through feature extraction- the selection the programmer makes about coding in the proper heuristics for decision-making. Deep learning circumvents the challenges of feature extraction. These models learn what is important by themselves via training. Training these models involves exposing the neural net to a large quantity of training examples. Afterwards, the model recalculates the input weights to maximize results and minimize errors.

TensorFlow- Google’s Open Source Software Library:

Deep Learning Algorithms

TensorFlow is an open source software library created by Google’s Brain Team. An open source software library is a collection of resources that are used to help computer scientists create programs and code. Google originally created it as a tool to train and build neural networks that decipher correlations and patterns that humans use. It is currently being used by Google as a replacement for Distbelief- Google’s first generation proprietary machine learning system. Companies that are already using TensorFlow include Google, Uber, Airbnb, Snapchat, Intel, and SAP. TensorFlow’s initial release was on November 9, 2015, and its first stable 1.0 version was released on February 5, 2017.

TensorFlow Compared to Other Libraries:

Deep Learning Algorithms

TensorFlow is quickly becoming the most used open source software library on the market. Since it is available via free download, it is easily accessible to individuals and firms alike. TensorFlow has two huge competitors: Theano and Torch. However, Google recently announced that it will shift over its Deep Mind division from Torch to TensorFlow. With such a large following from all different types of computer scientists, Google’s backing of TensorFlow will continually increase its usage.

Regarding performance, TensorFlow is clearly the best at presenting computational graph visualizations. Basically, TensorFlow provides a more aesthetically pleasing library than its competitors. TensorFlow also allows for model checkpoints. One can create, run, teach, and pause a model  for reevaluation. Therefore, model accuracy and precision increase, leading to faster creation. Additionally, the function TensorBoard allows the user to create a visualization of new models during building and debugging. In other software this feature may not be available, so those users need to mentally think of the context for new frameworks rather than seeing it on their screens.

Early Results from TensorFlow:

Deep Learning Algorithms

TensorFlow has already produced solid results for Google. Google’s first use of TensorFlow was to help with Google Translate. Google Translate was using the same translation algorithm since its inception in 2006, and technology had dramatically improved in those 10 years. Its speech recognition and image recognition improved, but the Google Team had trouble improving its machine translation. Google introduced the Google Neural Machine Translation system, teaching its machines to use the entire sentence as an input, rather than single words or phrases. Therefore, Google Translate now thinks more like a human and less like a machine.

Google is also developing a machine that can detect and predict diabetic retinopathy, a leading cause of blindness in adults. The machine can analyze retinal photos of patients and discern whether they have this form of blindness. According the Journal of American Medical Association, Google’s machine performed correct diagnoses at the same rate as human ophthalmologists. Google’s machines are also as accurate as dermatologists in identifying certain types of skin cancer. Google’s hope is not to cut out human doctors, but rather screen as many people as possible to prevent these diseases from worsening.

As deep learning improves, TensorFlow will play a more vital role in artificial intelligence. TensorFlow’s technology will continue helping users design and promote their work at a quicker and more efficient pace.

I Know First’s Development In Machine Learning 

Deep Learning Algorithm

The system is a predictive stock forecast algorithm based on Artificial Intelligence and Machine Learning with elements of Artificial Neural Networks and Genetic Algorithms incorporated in it.

This means the algorithm is able to create, modify, and delete relationships between different financial assets. Based on the relationships and the latest market data, the algorithm calculates its forecasts. Since the algorithm learns from its previous forecasts and is continuously adapting the relationships, it adapts quickly to changing market situations.

Deep Learning Algorithm

The I Know First Market Prediction System models and predicts the flow of money between the markets. It separates the predictable information from any “random noise”. It then creates a model that projects the future trajectory of the given market in the multidimensional space of other markets.

The system outputs the predicted trend as a number, positive or negative, along with the wave chart that predicts how the waves will overlap the trend. This helps the trader decide which direction to trade, at what point to enter the trade, and when to exit.

The model is 100% empirical, meaning it is based on historical data and not on any human derived assumptions. The human factor is only involved in building the mathematical framework and initially presenting to the system the “starting set” of inputs and outputs.

From that point onward, the computer algorithms take over, constantly proposing “theories”, testing them on years of market data, then validating them on the most recent data, which prevents over-fitting. If an input does not improve the model, it is “rejected” and another input can be substituted.

This bootstrapping system is self-learning, and thus live. The resulting formula is constantly evolving, as new daily data is added and a better machine-proposed “theory” is found.

Some stocks are members of several separate modules. Thus, multiple predictions can be obtained based on different data sets. Also, each module consists of a number of sub-modules, each giving an independent prediction. If sub-modules give contradictory predictions, this should be a warning sign. Six different filters are also employed to refine the predictions.


Deep Learning Algorithms

Predictions are generated daily for a growing universe of over 7,000 securities, including stocks, world indices, ETFs, interest rates and more for the short, medium and long-term horizons. The algorithm is applied to discover the best investment opportunities and used as a decision support system for existing investment processes or to develop systematic trading strategies. It is adaptable and scalable allowing comprehensive customized algorithmic solutions including integration of additional markets depending on clients’ needs – family offices, wealth management firms and hedge funds – as well as fund management partnerships. Furthermore, by offering top-notch technology to retail clients, I Know First also empowers private investors to identify opportunities in markets and manage their portfolios with more confidence. To read more about I Know First click here.