Real-time analytics is set to gain exponential traction during the next five years. When organizations envision capabilities like marketing 4.0 or supply chain 4.0, real-time visibility and analytics capabilities become imperative. To put it crudely, you can never build Industry 4.0 capabilities without perfecting real-time analytics.

And to perfect real-time analytics, you must perfect your AI at the edge strategy, architecture, and implementation. Consider your warehousing operations as an example. Have you developed a high-level architecture of areas within your warehouse where you need TinyML enabled TinyML-enabled IoT devices? If not, you need to hurry up. Similarly, have you decided how to leverage AI at the edge of your retail floors to capture and analyze real-time demand and customer behavior?

Though multiple algorithms can be leveraged on the edge, deep learning carries the most potential. If designed and implemented correctly, the opportunities you can create are plenty. This article will review some deep learning algorithms for AI at the edge. Please note that the article assumes familiarity with the basics of deep learning.

Fully connected models

Fully connected deep learning models are the simplest type of deep learning model. These models consist of layers of neurons that are stacked. The input data is fed directly as a long series of numbers for a fully connected model. These models can effectively learn most functions. However, their drawback is their struggle with spatial relationships in their inputs (for example, which values in the input are next to one another).

In embedded AI, fully connected models work effectively for discrete values (for example, if the input features are a set of statistics about a time series). Still, they aren’t as great with raw time series or image data. These models are well supported on embedded devices, with hardware and software optimizations commonly available.

Convolutional models

Convolutional models address the spatial drawback of fully connected models. Convolutional models can leverage spatial information as their inputs. These models are hence leveraged prolifically in edge devices. Examples of applications are recognizing shapes in images or identifying signal patterns in time series sensor data. This algorithm is widely leveraged and supported because spatial information is critical in signals and streaming data captured by devices like sensors.

Sequence models

Sequence models, like time series signals or written language, were initially designed for data sequences. To help them recognize long-term patterns in time series, they often include some internal “memory.”

The growing consensus among data science experts is that these models are very flexible and effective on any signal where spatial information is critical. While convolutional models are still the most popular regarding AI at the edge, there is a general belief that these models will eventually surpass convolutional models in applications of AI on the edge.

A current challenge with these models is that they are less supported on embedded devices than convolutional and fully connected models. There are comparatively very few open-source libraries that provide optimized implementations. However, this is not due to technical limitations. Hence, as this type of model’s popularity increases, support will also increase.

Embedding models

An embedding model is a deep learning model designed and pre-trained for dimensionality reduction. As evident with the specialization of these models, they can ingest a large number of dimensions and build a representative smaller set of numbers. They are leveraged similarly to signal processing algorithms. They primarily produce features that other models can use.

A widespread application of embedding models is for transfer learning. Transfer learning is an approach that reduces the amount of data required to train a model. Note that embedding models can be fully connected, convolutional, or sequence models. Hence the support varies, as discussed in previous sections, for each model type.

As mentioned previously, how to leverage which algorithm, on which devices, and where- are examples of elements that you will need to strategize. This is what will be one of the foundational ingredients for real-time analytics success.

References:

• Real-Time Analytics: Techniques to Analyze and Visualize Streaming Data – Byron Ellis
• Unlocking The Value of Real-Time Analytics- Christopher Gardener
• Building Real-Time Analytics Applications: Darin Briskman