A Gentle Introduction to Generative Adversarial Networks (GANs) with Marketing Applications
Have You Ever Found Yourself in the Following Situations?
- You have to list your products online but don’t know how to write an eye-catching description.
- You don’t have enough photographs of the product to attract viewers.
- You only have low-quality images that don’t immediately communicate the value of your product.
In such cases, the chances of your product being sold out are low, as engagement towards any product totally depends on how people perceive the information provided by you, and if it’s not top-notch then it automatically reduces your chance.
In the marketing world, there is a common saying that the product description is equally important as the product itself and anyone will only consider the listed product if they are convinced by its description and images.
In the real world, many listers often find themselves in a similar situation, where writing the content and preparing images are big tasks. This torments them as they don’t know what exactly needs to be done to attract viewers.
So is there any solution to the above scenario which can ease the user experience and can automate the complete process?
The answer is YES, you can, by using the power of Generative-Adversarial Network also popularly known as GAN.
Understand With An Example
Consider the scenario of an online marketplace for buying and selling products to see how GANs can be implemented.
GAN models can automate the creation of product listing descriptions. GANs understand and replicate the text patterns from successful product listings and help new listers with the description based on the product image or name, thus increasing the chances of their product being sold out quickly.
It can also help in generating different product angles by using a single frame photograph along with enhancing the quality of poor pixel images. With this, photographers can focus on taking one great picture, saving them the time, cost, and energy needed to take different angles of the same product.
How Is It Possible?
GAN learns from real-life historical data by the neural network approach.
Based on the dataset, consisting of images with descriptions, it tries to learn those parameters that impacted the performance and were able to engage more customers in the past.
The model tries to predict things based on the random noisy data as input and then counters them by comparing it with the genuine data points. This constant feedback during training helps it to learn and improve the understanding of the problem.
By the end of the training step, it understands the features that impacted the decision-making in the past and becomes self-sufficient in generating the data with decreased variation from given historical data and learns to mimic it on its own for any new data entries.
It’s a framework that is ideal for evaluating and producing a new creation. That is why it is also referred to as the imaginative and innovative side of artificial intelligence.
Diving Into GAN Architecture
(Technical discussion ahead. You can skip to the next section for application.)
GAN (Generative-Adversarial Network) is a class of artificial intelligence that works by setting two neural networks against one another — hence, adversarial.
This approach helps the architecture in learning the structure and the representation of complex real-world data, which can further be used to improve the AI pipelines, find anomalies, secure data, and generate similar synthetic examples.
The concept of Generative Adversarial Networks was first introduced by Goodfellow et al. in their 2014 paper “Generative Adversarial Networks” [link] with an unsupervised approach.
They proposed a training setup, where two neural networks, generator and discriminator, are pitched against each other in a competition. Both networks’ architecture consist of multiple convolutional layers, batch normalization, and ReLU with skip connections.
The generator network learns to create synthetic output and the discriminator networks try to understand and differentiate which is fake and which is actual, evaluating them for authenticity. Each side learns the methods of the other in a constant improvement cycle.
A discriminative model learns only the general threshold to differentiate between the categories and is not able to describe the category themselves. The model learns conditional probability.
A generative model learns about the individual categories and how the values are distributed in each one of them. It can be used to classify whether a certain value belongs to one category or another. This model learns joint probability distribution.
In further developments, many people experimented with GANs by providing them with some prior knowledge, commonly called “conditional” approaches.
Some examples of such modified GANs are – Pix2Pix network, CycleGAN architectures, etc.
- Pix2Pix- Random sketch gets auto-completed to the final image.
- CycleGAN – Famous example of creating zebra from horse image.
GAN has huge potential when implemented in real-world scenarios as it can learn and mimic the data in any domain: speech, text, image, human gesture and emotion, and various others.
It’s pretty clear how powerful this technique is. And there are many industries already leveraging its benefits. But what other areas are still possible with regards to its implementation? Let’s look into these questions in the next section.
Real-Life Applications of GAN
- In recent times the advertising and marketing industries have had the highest GAN adoption rate. This field has a huge scope of GANs implementation, as there is always a demand for unique but repetitive content (such as capturing images of photo models) to promote a certain product or service.
Face customization (Generative Photos).
Left to Right: original, adding smile and age, race editing, hair color editing
- GANs not only generate artificial photos but can also animate them and make them live, adding expressions, voice, and more to the photo.
- Recently, Cadbury implemented this in a similar fashion using Shahrukh Khan’s voice and face along with the expression to promote local shops during Diwali season. Link
- Many mobile applications are also using GANs such as FaceAPP, Prizma, etc.
- As stated in the famous quote, “Data is the new oil.” Keeping this in mind, many companies have nowadays started adopting GANs to produce new data in “data-limited situations” as it can understand real-world scenarios and can be applied in many domains.
- Implementation in chatbots is also a good example as responses from GANs appear more natural, giving a human touch to the conversation, or by adding emotions to the text when a query is made.
- Amazon has also developed the ec2GAN (e-commerce-conditional-GAN) for predicting the demographics of the customer, selling price, and the general dates for sales.
- In the advertising and marketing sector, GANs are a very powerful tool used for personalizing the user experience. We can consider an example where we want to customize any product based on the image provided.
Let us consider the below image as our input to the products that we want to customize.
Below are the results in which the above image is used to customize the headphone and phone cover.
You can provide any image and get your customized product ready.
This is one way of personalizing the user experience.
- Facebook also implemented GANs but with little modification and named them as IC-GAN. It’s a new image generation model which is capable of producing high-quality, realistic images, even if it is not present in the training dataset. It is also capable of generating image combinations that will rarely exist — for example, a penguin in the desert or kangaroos jumping around the city.
Other Areas Where GAN Can Be Implemented:
- GANs for recommendation: With the help of GANs we can generate possible combinations of products that are usually bought together. But as the variety of products in the market is changing very dynamically we need to keep our combinations updated based on trends. New products might not be in the list but can be a very good recommendation for a “products to buy together” category. This problem can be tackled with the implementation of GANs as it learns the pattern change going on in the market and incorporates it into the existing recommendation model.
- In the medical domain, not many images are present in order to train the network and are incredibly expensive to annotate. GANs can be helpful for synthesizing familiar images and training the model for further implementation such as detecting the disease just by looking at images.
- Another example of an application in the medical domain is using GANs to denoise images from CT and MRI, thereby reducing the need for high definition images and directly reducing the use of less powerful equipment while lowering harmful radiation from that equipment.
- Producing any sort of convincing text or image, just by giving it a set of parameters, which will be helpful in the e-commerce industry particularly for creating marketing campaign assets and descriptions.
- GANs can customize the outfits of a fashion model and generate different poses. They can also create a virtual model that resembles the target audience and fits into the brand’s image.
- It can also be used to create entirely new clothing color palettes or to recreate a three-dimensional model of an object from its two-dimensional images.
- It can generate marketing texts and reword the phrasing of a product description to include user-defined keywords. Texts produced by GANs are more or less similar to human-written content. It can also convert text to images, thus generating examples of products matching its textual descriptions.
These are just a few examples, it can be implemented in various other industries depending on the use case and provide value to new age business models.
Things to Consider
The key challenges with GANs currently are with its scalability, the computational cost to train, high sensitivity towards hyperparameter selections, and more.
But it’s just a matter of time before these things are resolved since GANs are currently the hottest topic in the data science field, going through constant improvements with the new research.
GANs are still in their early stage of implementation and there is a lot to explore and implement. But in the future, this will be one of the key algorithms that will be responsible for shaping the working of many industries and transforming currently used business models.