Zero-shot learning and the foundations of generative AI

We can remember 2022 as the year that cognitive AI techniques moved from the labs into widespread use. ChatGPT, a conversational AI that answers questions, went from zero to one million users in less than a week. The DALL-E 2, Midjourney, and Stable Diffusion imaging AIs opened up public access and captured the world’s attention with the variety and quality of images generated from phrases and short sentences.

I’ll admit, I had a little fun with DALL-E 2. Here’s his take on two lost souls swimming in a fish tank, and Tim Burton describes the agony of opening a green avocado.

“AI has generated headlines for projects like autonomous vehicles like Tesla and Waymo, unbeatable games (think AlphaGo), and a captivating generation of art like DALL-E,” says Torsten Grabs, director of product management at Snowflake.

Many machine learning models use supervised learning techniques in which a neural network or other model is trained using labeled data sets. For example, you can start with a database of images labeled cats, dogs, and other pets, and train a CNN (convolutional neural network) to classify them.

In the real world, labeling data sets at scale is expensive and complex. Healthcare, manufacturing, and other industries have many disparate use cases for making accurate predictions. Synthetic data can help augment data sets, but supervised learning models are still expensive to train and maintain.

One-shot and zero-shot learning techniques

To understand generative AI, start by understanding learning algorithms that don’t rely on labeled data sets. One-shot and zero-shot learning algorithms are examples of approaches that form the basis of generative AI techniques.

This is how ChatGPT defines one-shot and zero-shot learning:

“One-shot and zero-shot learning are techniques that allow models to learn and classify new examples with limited amounts of training data. In one-time machine learning, the model is trained on a small number of examples and is expected to generalize to new, unseen examples that are drawn from the same distribution. Zero-shot learning refers to the ability of a model to classify new, unseen examples that belong to classes that were not present in the training data.

David Talby, CTO of John Snow Labs, says, “As the name implies, one-shot learning aims to classify objects from one or just a few examples. The goal is for humans to prompt a model in plain language to successfully identify an image, phrase, or text.”

One-shot learning is done with a single training sample for each sample, for example, a headshot of a new employee. The model can then calculate a similarity score between two photos of faces, such as a photo of the person compared to the sample, and the score determines a sufficient match to grant access. A unique learning example uses the Omniglot dataset, a collection of 1,623 hand-drawn characters from 50 different alphabets.

In zero-trigger learning, the network is trained on images and associated data, including captions and other contextual metadata. An approach to zero-shot learning uses OpenAI’s CLIP (Contrastive Language Image Pretraining) to reduce the dimensionality of images in encodings, create a list of all possible tags from the text, and then calculate a matching similarity score. with the image of the label. The model can then be used to classify new images into tags using a similarity score.

OpenAI’s generative AI DALL-E uses CLIP and GAN (generative antagonistic networks) to perform the inverse function and create images from text.

Few shot learning techniques applications

One application of few shot learning techniques is in healthcare, where medical images with their diagnoses can be used to develop a classification model. “Different hospitals may diagnose conditions differently,” says Talby. “With single or multiple shot learning, the clinician can request algorithms, without using code, to achieve a certain result.”

But don’t expect fully automated X-ray diagnostics too soon. Talby says, “While the ability to extract information automatically is very valuable, learning from one, a few, or even zero tries isn’t going to replace medical professionals any time soon.”

Pandurang Kamat, CTO of Persistent, shares other potential applications. “Zero-shot and low-shot learning techniques open up opportunities in areas such as drug discovery, molecule discovery, zero-day vulnerabilities, case deflection for customer service teams, and others where labeled training data can be difficult.”

Kamat also warns about current limitations. “In computer vision, these techniques work well for image recognition, classification, and tracking, but can have problems in scenarios that require high precision, such as identifying cancer cells and marking their contours in pathological images.” , says.

Manufacturing also has potential applications for few-shot learning in defect identification. “No well-run factory will produce enough defects to have a large number of defect class images to train on, so algorithms must be built to identify them based on a few dozen samples,” says Arjun Chandar, CEO of IndustrialML.

Conceive next-generation AI solutions

Data scientists can try single-shot and zero-shot learning approaches to solve classification problems with unlabeled data sets. Some ways to learn the algorithms and tools include using Amazon SageMaker to create a news-based alert system, or using zero-shot learning in conversational agents.

Developers and data scientists should also consider new learning techniques and available models as building blocks for new applications and solutions rather than models optimized for specific problems. For example, Chang Liu, director of engineering at Moveworks, says developers can take advantage of large-scale NLP (natural language processing) models instead of building them themselves.

“With the introduction of large language models, teams are taking advantage of these intelligent systems to solve problems at scale. Instead of building an entirely new model, the language model only needs to be trained on the task description and appropriate responses,” says Liu.

Future AI solutions may resemble today’s software applications, with a mix of proprietary models, open source and commercial embedded components, and third-party services. “The achievements are within the reach of almost any company willing to spend time defining the AI ​​solutions problem and adopting new tools and practices to drive initial and continuous improvements,” says Grabs of Snowflake.

We are likely to see new AI learning approaches and achievements in 2023, so data science teams must continually research, learn, and experiment.