Deep neural networks and GPU computing have turned live video into structured, searchable, and actionable data, transforming how businesses use video analytics.
Valuable business insight requires robust, intelligent systems powered by extensive data. PwC research cited in the article projects the value of commercialised data to grow beyond $300 billion by the end of 2022. Combining machine learning with deep learning based on neural networks becomes increasingly accurate as more data feeds the systems, ultimately outperforming traditional methods despite requiring significant training effort.
Training Deep Neural Networks
Deep learning emerged in the mid-2000s as industries adopted it to train computers to mimic human capabilities, enabling systems to continuously enhance their sophistication and drive other AI applications. Neural networks require exposure to tagged data and large volumes of information to detect, identify, and classify objects in imagery or video. Unlike rule-based fraud detection, AI models can learn continuously from millions of daily transactions rather than waiting for humans to write new rules.
Compute, Hardware, and Software
Advances in GPU and cloud computing enabled processing of video at scale, and modern solutions leverage GPUs and deep learning to convert live video into structured data with rich metadata. The article highlights IBM’s Power System AC922 server paired with NVIDIA Tesla V100 GPUs, noting superhighway data transfers up to 5.6 times faster than before, with IBM’s Power AI framework easing data preparation and deep learning management for data scientists.
How the Models Learn
- Activation functions apply nonlinearity to output as data moves between stacked layers of perceptron units performing matrix multiplications.
- Supervised methods extract patterns by comparing predictions against ground truth.
- The difference between ground truth and prediction is the classification error, which weights and parameters are optimised to minimise.
Recommendations for Integrating Video and Analytics
- Start with the broader business picture: identify specific, viable business cases before investing in video, networks, and bandwidth.
- Develop a robust privacy policy, since business use cases and content-privacy concerns vary.
- Use automation with human checkpoints, keeping human experts as the second point of review for accuracy and reliability.
Conclusion
AI and deep learning make video searchable, actionable, and quantifiable. Operators can review footage in minutes that once took hours with conventional methods, rapidly identifying objects of interest and extracting maximum value from their video data.