Technology

Out Technology

Axyon Platform

Finance-focused deep learning platform, allowing extremely fast model prototyping and development.

Axyon Genetics

Fully automated feature-selection tool powered by genetics algorithms and deep learning.

Axyon Kerberos

Deep learning pipeline that allows usage of asynchronous data for synchronous predictions with minimum information loss.

Technology Highlights

  • Axyon Platform

    The Axyon Platform is a finance-focused deep learning platform that allows extremely fast model prototyping and development.

    To give context, it’s important to note that the classical approach to deep learning is to use low-level programming libraries to manually code predictive models that then have to be trained, evaluated and tested over never-seen-data. The problem lies in the fact that, while deep learning models easily outperform traditional statistics, they are characterized by a high level of hyperparameters (e.g. network morphology) that require lengthy fine tuning and tests. For this reason, deep learning model development, when done manually, requires a large amount of time. Some deep learning platform are on the market (both free and commercial) but they are focused on artificial vision and image processing (e.g. Nvidia Digits).

    The Axyon Platform changes things by providing a web-based interface that allows an extremely fast development and refining process for deep learning models. For example, several tests of concurrent morphologies can be launched with few clicks, cutting development time and improving performances. Additional tools such as dataset visualization, training process stats and tuning, and evaluation metrics are specifically designed for numeric problems, making the Axyon Platform tailored for finance prediction problems.

  • Axyon Genetics

    Axyon Genetics is a fully automated feature-selection tool powered by genetics algorithms, that helps in reducing the noise of difficult dataset (peculiar of difficult problems such as trading-related predictions). Included in the Axyon Platform’s user interface, automatically evolves DNAs representing enabled features over all the ones present in the dataset. Several policies can be used and customised to tailor problem-specific fitness functions.
  • Axyon Kerberos

    Kerberos is a Deep Learning pipeline that allows using asynchronous data for synchronous predictions with minimum information loss.

    In classic machine-learning approaches, dealing with asynchronous data is a problem when looking for synchronous predictions, because a manual “flattening” step is in this case required to compress asynchronous data belonging to the same synchronous label in a single feature vector.

    In the specific cases where data order is important inside the same synchronous step, recurrent network approaches can be used, although with some difficulties; but, in the more common situations where data order is not important (or too uncertain to be important) inside the same step, the recurrent network approach is detrimental.

    Kerberos innovates this scenario by using convolutional networks techniques (typical of image processing) for giving away with data order, and to naturally (and automatically) compress the asynchronous event in synchronous “snapshots” that represent the information conveyed by the asynchronous data without manually-induced information loss.

    Some use cases:

    • Synchronous predictions with asynchronous data: data is available in an asynchronous manner (e.g. a variable number of events occurring each day) but the required prediction is synchronous (one prediction per day).
    • Synchronous predictions with asynchronous/synchronous data mix: data is available from different sources, both asynchronous and synchronous (e.g. a variable number of financial news for each day and a stock’s fixed-length market indicators feature vector per day) but the required prediction is synchronous (one trading signal per day).
    • Sample classification with asynchronous/synchronous data mix: data is available from different sources, both asynchronous and synchronous (e.g. a user’s activity on an ecommerce website combined to their demographic data) and a sample classification prediction is required (whether they will complete a purchase).