Data science /

Connecting the dots in Neural Networks

February 7 2021/3 min read

There are countless resources on the internet explaining what Neural Networks are and how they model the data to extract predictions. Theoretical definitions are great as they give you the foundations to understand how things work. However, when you want to put your newly acquired knowledge into practice, the availability of easy-to-use tools decreases substantially, especially if you prefer to avoid programming.

In this blog, I provide you with a quick guide on how you can connect neural network concepts and put them into practice by using Peltarion's AI no-code platform!

02/ First, what is Peltarion's AI platform?

The Peltarion platform is an operational AI platform for building and deploying AI models that is capable of executing a broad range of deep learning-based use cases. The platform carries prebuilt and pretrained networks ready for the user to work on, as well as provides tools to create own architectures. The platform focuses on ease of use for iterating, evaluating and refining models.

03/ What is a neural network, again?

I’m sure you have read many times definitions of neural networks, I’ll provide you with a short one so we are in consensus with some terminology. 

Neural Networks are non-linear modeling tools used to find patterns in data and the relationship between inputs and outputs. The input refers to the data point we have and from which we want a neural network to be able to extract or predict something: the output or target. For example, given an image of an impeller (input) we want a neural network to predict if the image represents a defective or non-defective part (target).

Neural network used for image classification, inference from our tutorial Detecting defects in mass-produced parts

Similarly, given a text extract, a neural network can predict the topic of the text (target). In our example below the topic “History” was predicted from the input text.


Neural network predicted “History” as the book genre. Inference from our tutorial Classify text in any language

04/ How do neural networks work?

A neural network is composed of a large number of processing units called neurons or nodes that are highly interconnected with each other and arranged in structures called layers. Between the input layer and output/target layer typically are other layers called the hidden layers. If there are many hidden layers, the network is referred to as a “deep” neural network. The nodes on each layer receive the input from nodes on the previous layer and pass their output to the next layer. 

The hidden layers

The hidden layers apply transformations to the data one after the other. On the Peltarion platform, data transformations are performed by basic building units called blocks. For input data coming from images, the 2D Convolution block is a common block to use and when working with any kind of numerical data, the most common block used as a hidden layer is the Dense block.

A (convolutional) neural network architecture and its implementation on the Peltarion platform

Each block has configurable parameters depending on the type of input or target data and the transformation they are executing. For instance, having images as input, you can select custom values for image augmentation and on the 2D Convolution block, you can select the number of filters, filter's width and height, horizontal and vertical stride, activation function, padding, etc.

For text as input, the multilingual BERT (Bidirectional Encoder Representations from Transformers) and the Universal Sentence Encoder (USE) snippets have hidden layers specialized in processing text data in multiple languages that are not accessible as blocks but instead as (pre-trained) snippets on the Peltarion platform.

05/ Snippets: gateway to neural networks architectures and pretrained models

Many of the most powerful neural networks have very large architectures which can make them tedious to build block by block. On the Peltarion platform, we’ve prebuilt popular networks as snippets so you don’t need to build them yourself e.g. EfficentNet snippet.

Most of the snippets available are even pretrained, for example English BERT, multilingual BERT or XLM-R ← SBERT-nli-stsb called XLM-R Embedding on the platform so you can take advantage of transfer learning. When using transfer learning, the knowledge acquired by these models from vast amounts of data can be transferred to solve other, related problems in your field.

Available neural networks snippets and pre-trained models on the modeling view of the Peltarion Platform.

06/ Why using neural networks?

A neural network learns to interpolate data and that is why you train it by giving both the input and the correct target. Their power relies on their exceptional capacity to adapt and learn very quickly, extracting predictions as a result of a multiple series of activations and transformations executed on each node and on each layer. Here are three key advantages of using neural networks:

  1. As mentioned earlier, neural networks are non-linear which makes them extremely versatile. Instead of deciding if you need to interpolate with a polynomial, Fourier series, or with an exponential function, you use neural networks to have good chances of finding the complex relationships in the data.
  2. Neural networks learn patterns in the data. To achieve good interpolation, they have to encode the input in an efficient numerical way. You can also use this encoding for side-applications like similarity search.
  3. Neural networks are great at working with problems in high dimensional space. For example: 
  • The defect detection classification problem is using an image of size 256 x 256 x 3 pixel values which equals to about 200.000 dimensions to start with on the first layer.
  • An English paragraph as an input for the book genre classification example can contain 150-200 tokens, with approx. 110k possible tokens on the multilingual BERT model’s vocabulary, the dimensions go up to several million (16M - 22M).
  • Many input parameters or the combination of different types of input data is also considered as high dimensionality. For instance, combining text or audio with tabular data or images and tabular data as in our Predict real estate prices tutorial. There we are combining map images from Open street map and tabular demographic data as median house age, total rooms, total bedrooms, population, households and median income.

07/ Try yourself!

Finally, it is time to roll up your sleeves and get some hands-on practice! Build some powerful AI solutions with neural networks on the no-code Peltarion platform, for some inspiration check out our tutorial catalog. 

If you want to explore more in detail how to build your own project or need some guidance, please get in touch with me at liliana@peltarion.com. We’d love to hear from you!

  • Liliana Lindberg

    Liliana Lindberg

    Solutions Architect

    Liliana works as a solutions architect at Peltarion guiding customers to solve business challenges by using AI. She is passionate about emerging technologies and before joining Peltarion she worked for a number of years at Google as a GCP customer engineer. Her academic background includes BSc. Systems and Computing Engineering; MSc. in Geographical Information Systems from the University of Calgary, Canada; and a Master’s level Business Leadership Specialization from Duke University.

02/ More on Data science

Contact us