What is deep learning? | McKinsey

deep learning Although it's been around for a while, most of us had never used deep learning-based tools until the release of OpenAI's ChatGPT in late 2022. (And while we were amazed by ChatGPT's output, most of us didn't even know it was being used.) Deep learning to generate them. ) Like its predecessor DALL-E, Google's Imagen and PaLM, and Stable Diffusion, ChatGPT Use large-scale deep learning models trained on large datasets to generate content based on prompts. But unlike previous versions, ChatGPT operates through an open access API, allowing the general public to experience the power of deep learning for the first time.

The world of artificial intelligence and machine learning (with deep learning being the next evolutionary step) is undergoing a generational shift from ideas studied by scientists to tools used by all kinds of people for all kinds of tasks. . McKinsey analysis shows that between 2015 and his 2021, the cost of training his image classification systems (run on deep learning models) decreased by 64%. During the same period training time he improved by 94%. We also found that generative AI (gen AI) could bring up to $4.4 trillion worth of annual benefits to the global economy. All of these major changes are being made possible by deep learning.

But what actually is deep learning? And how is all this possible? Read on to find out.

Learn more about McKinsey Digital.

What is machine learning?

Before moving on to deep learning, let's understand the basics. Machine learning is a type of artificial intelligence that can adapt to a wide range of inputs, including large datasets and human instructions. These algorithms can learn to detect patterns and make predictions and recommendations by processing data and experience, rather than receiving explicit programming instructions. Algorithms also adapt in response to new data and experiences, improving over time.

The amount and complexity of data currently being generated is too vast for humans to comprehend, increasing the need for and potential for machine learning. In the years since its widespread adoption, machine learning has impacted many industries, including medical image analysis and high-resolution weather forecasting.

To learn more about machine learning, visit McKinsey Explainer.

How is deep learning different from machine learning?

Deep learning is a more advanced version of machine learning that is particularly good at handling a wide range of data resources (text, and unstructured data, including images), requires even less human intervention, and often , can produce more accurate results than traditional machines. learn. Deep learning uses neural networks, based on the way neurons interact in the human brain, to ingest and process data through multiple layers of neurons that recognize increasingly complex features of the data. For example, an early layer of neurons might recognize something as having a certain shape. Based on this knowledge, a later layer could potentially identify the shape as a stop sign. Similar to machine learning, deep learning uses iteration to self-correct and improve predictive capabilities. Once it “learns” what objects look like, it can begin to recognize objects in new images.

What is the relationship between deep learning and artificial intelligence?

ChatGPT makes AI visible and accessible to the general public for the first time. ChatGPT and other language models like it were trained with a deep learning tool called a transformer network to generate content in response to prompts. Transformer networks allow the Gen AI tool to weight different parts of the input sequence differently when making predictions. Transformer networks consisting of encoder and decoder layers allow Gen AI models to learn relationships and dependencies between words in a more flexible way compared to traditional machine learning and deep learning models. Masu. That is, the transformer network is trained on vast swaths of the Internet (e.g., all the traffic footage ever recorded and uploaded) rather than on a specific subset of data (e.g., a particular image of a stop sign). This is for the purpose of As explained further below, underlying models trained on transformer network architectures, such as OpenAI's ChatGPT and Google's BERT, apply what they learn from a specific task to a more generalized set of tasks, including content generation. Can be transferred to tasks. At this point, you can ask your model to create a video of a car passing a stop sign.

The foundation model can create content, but it doesn't know the difference between right and wrong, or even what is socially acceptable and what is not. When ChatGPT was first created, training required a large amount of human input. OpenAI has hired thousands of human workers around the world to hone its technology, including cleaning and labeling datasets, reviewing and labeling harmful content, and flagging it for removal. This human input is a big part of what made ChatGPT so revolutionary.

What types of neural networks are used in deep learning?

There are three types of artificial neural networks used in deep learning.

• Feedforward neural networks. First proposed in 1958, this simple neural network allows information to travel in only one direction. That is, it does not move backwards to be reanalyzed by the model, but only forwards from the input layer to the output layer of the model. This means that you can feed or input data into a model and “train” the model to predict something about different datasets. As just one example, feedforward neural networks are used in the banking industry and other industries to detect fraudulent financial transactions. Here's how it works: First, we train a model to predict whether a transaction is fraudulent based on the dataset we used to manually label the transaction as fraudulent. The model can then be used to predict whether newly received transactions are fraudulent and can be flagged for further investigation or blocked entirely.

• Convolutional Neural Network (CNN). A CNN is a type of feedforward neural network whose connections are inspired by the organization of the brain's visual cortex, the part of the brain that processes images. CNNs are therefore suitable for perceptual tasks, such as being able to identify bird or plant species based on a photo. Business use cases include diagnosing diseases from medical scans and detecting company logos on social media to manage brand reputation and identify potential joint marketing opportunities. It will be.

  Here's how it works:

• First, a CNN takes an image, such as the letter “A,” and processes it as a collection of pixels.
• The CNN identifies unique features (for example, the individual lines that make up the letter “A”) in a hidden layer.
• The CNN can classify another image as the letter “A” if it finds that the new image has the same unique features previously identified as making up the letter.

• Recurrent Neural Networks (RNN). RNNs are artificial neural networks, and their connections contain loops. That is, the model moves the data forward and loops backwards to run the previous layer again. RNNs are useful for predicting emotions and the end of sequences, such as large samples of text, audio, and images. This is possible because each input enters the model not only alone, but also in combination with previous inputs.

  Continuing with the banking example, RNNs can help detect fraudulent financial transactions just like feedforward neural networks, but in a more complex way. Feedforward neural networks can help predict whether one individual transaction is likely to be fraudulent, whereas recurrent neural networks can help predict whether one individual transaction is likely to be fraudulent, whereas recurrent neural networks can help predict whether one individual transaction is likely to be fraudulent, whereas recurrent neural networks can help predict whether one individual transaction is likely to be fraudulent. It can “learn” from and measure each transaction. for that person's entire record. You can do this in addition to using the general training of feedforward neural network models.

To learn more about deep learning, neural networks, and their use cases, check out our executive guide to AI. Learn more about McKinsey Digital.

What is the base model?

The underlying model is a transformer network architecture, a deep learning model trained on large amounts of unstructured, unlabeled data. The basic model can be used as is or adapted to a specific task through fine-tuning, allowing it to be used for a wide range of tasks. Fine-tuning involves a relatively short period of training on a labeled data set. Typically a much smaller data set than the data set on which the model was originally trained. This additional training allows the model to learn and adapt to nuances, terminology, and specific patterns in small data sets. Examples of fundamental models include DALL-E 2, GPT-4, and stable diffusion.

What is a large-scale language model?

Large-scale language models are a class of fundamental models that can handle large amounts of unstructured text. These models can learn relationships between words or word parts (also called tokens). This enables large-scale language models to generate natural language text and perform tasks such as summarization and knowledge extraction. Google's Gemini runs on a large language model called LaMDA.

Click here for details mckinsey digital.

Which areas will benefit from machine learning and deep learning?

McKinsey has collected more than 400 use cases for machine learning and deep learning across 19 industries and nine business functions. Based on our analysis, we believe nearly every industry can benefit from machine learning and deep learning. Here are some examples of use cases that span several sectors.

• Predictive maintenance. This use case is extremely important for any industry or business that relies on equipment. Rather than waiting until equipment breaks, businesses can use predictive maintenance to predict when maintenance will be needed, reducing potential downtime and lowering operating costs. Machine learning and deep learning have the ability to analyze large amounts of multifaceted data, increasing the accuracy of predictive maintenance. For example, AI practitioners can overlay data from new inputs, such as audio or image data, adding nuance to neural network analysis.
• Logistics optimization. Optimizing logistics using AI can reduce costs through real-time prediction and behavioral coaching. For example, AI can optimize the routing of delivery traffic, improve fuel efficiency, and reduce delivery times.
• customer service. AI technology in call centers can help create a more seamless experience and more efficient processing for customers. This technology does more than just understand what the caller is saying. Deep learning analysis of audio allows you to assess customer tone. If an automated calling service detects that a caller is upset, the system can reroute them to a human operator or manager.

Click here for details mckinsey digital. If you're interested in working at McKinsey, check out our deep learning job openings.

Articles referenced:

• “Tech Generational Moment with Generative AI: A CIO and CTO Guide,” July 11, 2023, Aamer Baig, Sven Blumberg, Eva Li, Douglas Merrill, Adi Pradhan, Megha Sinha, Alexander Sukharevsky, Stephen Xu
• “A new, faster machine learning flywheel for enterprises” March 10, 2023, Medha Bankhwal, Roger Roberts
• “Deep Learning in Product Design,” December 14, 2022, Mickael Brossard, Jacomo Corbo, Marie Klaeyle, Bill Wiseman
• “The Executive Guide to AI,” November 17, 2020, Michael Chui, Brian McCarthy, Vishnu Kamalnath
• “The Origins and Pioneers of Deep Learning,” May 8, 2018
• “Notes from the AI ​​Frontier: Applications and Value of Deep Learning,” April 17, 2018, Michael Chui, James Manyika, Mehdi Miremadi, Nicolaus Henke, Rita Chung, Pieter Nel, Sankalp Malhotra