Deep learning, an AI technique, has gained popularity due to increased data and computing power. It underpins various everyday applications like language translation, face-tagging, smart replies, and generative models. ChatGPT, the record-breaking AI chatbot, relies on a deep-learning model trained on vast internet data. DALL-E, Midjourney, and Stable Diffusion are deep-learning systems that generate images from text descriptions, showcasing the connection between images and text.
Deep learning and machine learning
Machine learning, a kind of artificial intelligence that trains computers to execute tasks based on experience, is a subset of deep learning. Machine learning algorithms "train" themselves by analyzing annotated samples, in contrast to traditional, rule-based AI systems.
Machine learning proves valuable in scenarios where rules are unclear and cannot be explicitly coded. For instance, in fraud detection, a machine-learning algorithm is trained using a dataset of bank transactions and their outcomes (legitimate or fraudulent). By analyzing patterns in the data, the algorithm develops a statistical representation of common characteristics for both types of transactions.
When presented with new transaction data, the algorithm applies its learned patterns to classify it as legitimate or fraudulent. The accuracy of the algorithm improves with high-quality data, adhering to the principle that more data enhances machine learning performance.
Machine learning excels in diverse tasks, with various algorithms specialized for specific problem domains.
Deep learning & neural networks
Traditional machine learning algorithms are good at solving many problems that rule-based computers have trouble with, but they struggle to handle soft data like pictures, videos, audio files, and unstructured text.
Developing a breast-cancer-prediction model through traditional machine-learning methods involves the collaboration of numerous experts, including domain specialists, programmers, and mathematicians, as mentioned by AI researcher and data scientist Jeremy Howard in the referenced video.
The process involves extensive feature engineering, where experts program the computer to identify known patterns within X-ray and MRI scans. Machine learning algorithms are then applied to the extracted features. Constructing such an AI model is a time-consuming endeavor that can span several years.
Deep-learning algorithms address similar problems by employing deep neural networks, which are software architectures inspired by the human brain (although distinct from biological neurons). Neural networks consist of multiple layers of variables that adapt to the characteristics of the training data. As a result, they acquire the ability to perform tasks like image classification and speech-to-text conversion.
Neural networks excel at autonomously identifying common patterns in unstructured data. When trained on a diverse set of images, a deep neural network discovers methods to extract relevant features. Each layer of the network specializes in detecting specific attributes like edges, corners, faces, and even intricate details such as eyeballs.
Neural networks were conceived in the 1950s but received limited attention due to demanding data and computing requirements. However, recent advancements in storage, data availability, and computing resources have propelled neural networks to the forefront of AI innovation. Their potential has been widely recognized, leading to widespread adoption in various applications.
Distinct deep-learning architectures specialize in different tasks. Convolutional neural networks (CNNs) excel in pattern detection for images, while recurrent neural networks (RNNs) process sequential data like speech, text, and music. Graph neural networks (GNNs) predict relationships in graph data, such as social networks and online transactions.
The transformer architecture, widely utilized in large language models (LLMs) like GPT-4 and ChatGPT, has gained immense popularity. Transformers are especially suitable for language-related tasks and can be trained on vast amounts of raw text data.
What Is it used for?
There are several domains where deep learning is helping computers tackle previously unsolvable problems:
- Computer Vision
- Voice and Speech Recognition
- Natural Language Processing and Generation
- Art Generation
How's the future of it?
The Turing Award, equivalent to a Nobel Prize in computer science, recognized deep learning pioneers in 2019. However, ongoing efforts aim to enhance deep learning.
Advancements include interpretable deep-learning models, neural networks with reduced training data requirements, and edge AI models that operate independently of extensive cloud resources.
Despite being at the forefront, deep learning is not the ultimate destination for AI. Continued evolution could lead to entirely new architectures and approaches within neural networks.