First on Chain – LSTM Neural Nets Coming to the GNY Machine Learning Wallet

If a data sequence is long enough then traditional feed-forward neural networks can end up missing out on vital information from earlier in the ML process. As an example, imagine you want to classify what kind of weather event is happening each day in a course of a year across the Paris metropolitan area. It is not clear how a traditional neural network could use its reasoning about January’s weather events in the region to inform its classification of July’s ones.

Data sets that are best suited for LSTM feature seasonality. LSTM is capable of capturing the patterns of both long-term seasonality such as a yearly pattern and short-term seasonality such as weekly patterns. Additionally, single major events can be accounted for as they will have an expanded time impact on a system. For example- where people would book more days of accommodation in order to attend a sports event. LSTM neural networks have the ability to triage the impact patterns from different categories of events.

So LSTM models are recurrent neural networks capable of learning long-term time series dependencies, specifically tuned for series data. The different gates inside LSTM boost its capability for capturing non-linear relationships for forecasting. This is crucial because causal factors generally have non-linear impact on demand and demand predictions. When these factors are used as part of the input variable, the LSTM neural networks learn the nonlinear forecasting relationships.

To best explain the power of LSTM neural networks the GNY team selected the following two hypothetical use cases put forth by our community.

Hypothetical Client #1 – Amazon Delivery Services.

The dilemma – How can Amazon delivery optimize efficiency of delivery services.

The solution – Use GNY’s LSTM neural net to better understand delivery requirements based on different annual, seasonal, and specific date driven patterns to optimize staffing, and delivery efficiency.

How it would work – The most recent three years of delivery and general conditions data would be taken for the training dataset, and one year of data will be used for the test set. The GNY LSTM would then predict future events based on different time-scale seasonalities. With these predictions Amazon could more efficiently scale staffing for delivery needs, plus optimize routes, vehicle types, and options to increase the efficiency of delivery services.

Hypothetical Client #2 – The Australian Department of Agriculture’s NWI Groundwater Information System.

The dilemma – How can data help the Australian authorities, business, and homeowners more effectively respond to drought conditions and national water needs.

The solution – Use GNY’s LSTM neural network to better understand the multiple systems that converge in ground water systems. These include weather patterns, domestic and industrial water usage, non-weather climate events (ie. wildfires) and the efficiency of the various water delivery systems for domestic and industrials parties. The LSTM could efficiently predict both the demands on water and the changing resources available to meet those needs. Extending these predictions into the future would allow the Department of Agriculture’s NWI to predict when shortages will occur and develop plans that can prepare individuals and businesses accordingly. Better prediction of different annual, and seasonal patterns would increase preparedness and extend the amount of time available to respond meaningfully to potentially life threatening challenges.

Let us know in our Telegram chat group what you think of our potential use cases for GNY’s LSTM Neural Network. Please feel free to add your ideas for other areas you believe will benefit from this on-chain contract.