Deep neural networks: The key to next-level AI technologies

IMAGE CREDIT:
Image credit
iStock

Deep neural networks: The key to next-level AI technologies

Deep neural networks: The key to next-level AI technologies

Subheading text
Deep learning technologies can teach artificial intelligence systems to solve even the most complex problems.
    • Author:
    • Author name
      Quantumrun Foresight
    • November 14, 2022

    Artiļ¬cial intelligence (AI) systems based on deep neural networks (DNNs) have seen success in various applications, most notably in computer vision and Natural Language Processing (NLP). Deep neural networks power versatile algorithms that simultaneously process different modalities (e.g., language, vision, and sensory), resulting in more informed decision-making.

    Deep neural network context

    Deep neural networks, also known as deep nets, are crucial machine learning (ML) components that employ complex mathematical modeling. These neural networks mimic the hundreds of nodes in the human brain that offer people logic, reasoning, and the ability to refer to related information. Similarly, DNNs allow mathematical models to have several layers of output that they can refer to for more accurate and logical classifications. This process is known as deep learning (DL), where algorithms evolve beyond the basic input/output protocols. 

    There are three types of DNNs, all supporting different functions. 

    1. Artificial neural networks (ANNs) can learn non-linear functions and complex relationships between inputs and output. 
    2. Convolution neural networks (CNNs) are primarily used in video and image processing. These networks are responsible for the proliferation of controversial deepfake content. 
    3. Finally, recurrent neural networks (RNNs) solve audio, text, and time-series data problems. 

    The most common use cases of DNNs are computer vision, speech recognition, creating virtual environments, and medical diagnosis. However, whereas the traditional training of DNNs is usually in one modality only (e.g., vision or speech), some tech companies are developing versatile multimodal nets that are more capable of understanding relationships among complex datasets.

    Disruptive impact

    Self-supervised learning through DNNs, in which computers learn by examining the world directly rather than being instructed via labeled pictures, text, audio, and other data sources, has fueled many significant AI advances. In 2022, Meta announced its new self-supervised multimodal algorithm called data2ve. The algorithm works separately on speech, pictures, and text, outperforming other single-purpose computer vision and speech algorithms. According to the company, self-supervision allows computers to learn about the world simply by watching it and then determining the image, speech, or text structures. Machines that don't need to be explicitly taught to identify pictures or comprehend spoken language are far more scalable. Meta hopes this type of deep learning will allow for more adaptable AI.

    In particular, the highly evolved computing capabilities of DNNs are being tested to predict potential terrorist attacks. The use of ML methods in the study of terrorism is becoming more common because a massive amount of labeled data is available. Additionally, the improvements in computer hardware have allowed DNNs to create increasingly powerful computing systems capable of performing the required calculations. Deep neural networks can analyze several terrorist attack components. For example, the model can classify the general types of weapons used, the targeted region, and the kind of attack (e.g., suicide bombers and public shootings). These forecasts are crucial for counterterrorism. Deep learning may quickly generate these predictions, allowing police forces to devise procedures to prevent the loss of human life and infrastructure, including essential services.

    Implications of deep neural networks

    Wider implications of DNNs may include: 

    • The automotive industry using DNNs to improve self-driving technology to process different stimuli simultaneously, such as noises and obstructions.
    • Increasing intelligent chatbots and digital assistants that may perform accurate sentiment analysis.
    • The ability to forecast better for extreme weather and disasters. This ability may help warn people of potential storms, hurricanes, and forest fires.
    • Creating more sophisticated deepfake content for propaganda and creating disinformation which may increase global political instabilities.
    • Deep neural networks used to shape and test military strategies, such as brainstorming different scenarios and predicting their individual probability of success. 

    Questions to comment on

    • What are the other potential use cases of DNNs?
    • How might DNNs be abused in everyday applications?

    Insight references

    The following popular and institutional links were referenced for this insight: