While there may be a lot of hype about AI, there are many new and exciting deep learning applications with many practical applications.
Deep learning helps machines understand and generate “sensory data,” i.e. images, video, music, sound, text, and speech. The main areas that shine are around visual perception and natural language understanding. If you have problems that can be solved with computer vision or natural language processing, they can be solved with deep learning. Home visual searches are a good example — think AirBnB and HomeLuv — as well as question-and-answering services and voice assistants, such as Alexa and Google Duplex.
As the field of AI rapidly expands in its potentialities, online retailers are only scratching the surface of deep learning capabilities.
Some major players, such as Walmart, are using AI for product recommendations, but the store, just like many of its competitors, is not yet using AI to its full potential. Not because these retailers don’t have the teams or the capacity, they are simply not aware of everything that it’s possible.
Why? Because scientists don’t know about solving problems they don’t face, and businesses (as well as start-ups) might be ignoring difficult problems because they don’t know the research to solve them exists.
This lack of dialogue between scientists and business people often silos deep learning research to the domain of researchers and more often than not inadvertently excludes the businesses that this deep learning could best help.
Deep learning researchers need to spend time facing customer problems, and customer-facing experts should study deep learning. There should also be close collaboration between the two groups to create new practical applications to solve the biggest challenges.
It seems that artificial intelligence legend, Andrew Ng, agrees with me. He recently announced an upcoming course “AI for Everyone”. He says: “The AI-powered future must be built by both engineers and application domain experts. We will need millions of AI engineers. We will also need millions of experts from every industry to understand how to apply AI within their organizations.”
Before we dive into practical applications, it is important to understand how deep learning works.
“The information bottleneck” is a theory that best explains deep learning success. Deep learning models take “noisy data” as input, say an image, a sound, text, etc. and during the training phase, the information is reduced to what is essential to complete the task learned during training. For example image classification, object detection, language translation, etc.
A popular construct in deep learning is the autoencoder, which is an artificial neural network architecture that allows you to see the information bottleneck in action.
Original from: Image reconstruction
A source image is encoded/compressed and then decoded/decompressed. During the decoding, it is possible to add color to black and white images and also do other sorts of improvements like making a photo look like a painting.
Original from: Image captioning
A source image is encoded/compressed and then decoded/decompressed into a text sequence. How is this possible? Both text, and images in computers are represented as numbers. The decoder translates the numbers into text using a dictionary.
One simple way to think about an autoencoder (or AE), is that it compresses/encodes and decompresses/decodes input data to transform it from one format to another. It’s a neural-network architecture with many useful applications.
For example, AEs are useful in image transformation tasks, document cleaning, applying color to images, medical image segmentation, etc. When used for natural language processing they are also efficient in machine translation, document clustering, sentiment analysis, and paraphrase detection. AEs can even translate images, generate music, or produce video.
There is this myth that only companies like Google, Facebook and Amazon have enough data to make deep learning useful.
The truth is, however, that you can train deep learning models with very little data thanks to a standard technique called transfer learning. This technique allows you to start off your model from pretrained data on a closely related task.
For instance, Google’s AI lab was able to create a deep learning machine that “can assist doctors in decision making and improve efficiency in interpretation [of x-rays]” with minimal data sets (880 images with annotations). It is very difficult to get significant training data in this domain.
There are many publicly available datasets like ImageNet, Coco dataset, Stanford Q&A corpus, etc. You can train your models using these datasets for free using Google’s Colaboratory. Alternatively, you can start with pre-trained models from the Tensorflow Hub.
Here are some practical applications of Deep Learning that are often overlooked:
To revisit the earlier example of Walmart, Ajinka More, the former Principal Data Scientist at Walmart Labs, notes that there are many challenges in creating effective deep learning machines that can optimize product searches on Walmart’s website. He notes that “it’s hard to obtain a large amount of labeled data for this problem.
To circumvent this, we trained the title similarity model entirely on synthetic labels.” While Walmart can afford to assemble a data scientist team to create this expansive neural network, smaller retailers can too because a lot of the building blocks are public and free.
Ultimately, AI needs to be de-mystified. It’s the driver behind better, more relevant navigation and searches within a website and there are many untapped possibilities that could result in more robust sales and brand allegiance.
Hamlet Batista is CEO of RankSense.
The post Deep learning for online retailers: 5 practical applications appeared first on ClickZ.Reblogged 1 year ago from www.clickz.com