Overview

NVIDIA’s CUDA architecture — launched in 2006 — enabled developers to harness the parallel computing power of GPUs for more than just graphics. This laid the foundation for AI breakthroughs, allowing massive neural networks to train at speeds never seen before.

A variety of threat models have emerged as quantum computing continues to mature. They include threats to cryptographic systems utilized by the finance and e-commerce sectors, sampling speedups for machine-learning problems, risks of bias in training datasets or models, compromises of data integrity, and privacy concerns relative to sensitive data-sharing arrangements. Individuals want guarantees that their information is represented honestly in the model, even when the model is in the possession of those who control it. A novel approach to enhanced privacy-preserving techniques, brought to market using federated learning, addresses these concerns directly. Furthermore, as computing shifts to models running at the edge, the orchestration of cloud and edge becomes critical to long-term energy requirements. A decay in performance improvement from classical chip scaling and an increasing share of energy consumption from AI data centers heighten concerns about power consumption over time.

Budget limitations, infrastructure availability, bandwidth constraints, and real-time requirements drive much of the intelligence toward the edge of the network—that is, the vicinity of where the data is generated. These edge devices sense, act, and infer based on learned models that meet the energy budgets, static latency thresholds, and increasing workload support requests. Productive orchestration of the resulting distributed computational resources uses federated learning, where aggregates of model updates are collected and combined rather than the raw training data, often enabling a degree of differential privacy.

Classical data transfers are reduced, improving operational efficiency and reducing costs by constraining burdens on the core and regional fabric of the cloud. Such constraints also create opportunities to offer confidentiality to sensitive data, especially in regulated regimes. Nevertheless, privacy-preserving operational models add complexity to model deployment, hardware infrastructure needs to support privacy-preserving inference using the outputs of the federated learning process, and privacy-preserving inference models often require elevated processing capability. Availability of hardware accelerators at the predicted required price points and levels of performance will therefore strongly influence the commercial adoption of these models.

Key Insights

A variety of threat models have emerged as quantum computing continues to mature. They include threats to cryptographic systems utilized by the finance and e-commerce sectors, sampling speedups for machine-learning problems, risks of bias in training datasets or models, compromises of data integrity, and privacy concerns relative to sensitive data-sharing arrangements. Individuals want guarantees that their information is represented honestly in the model, even when the model is in the possession of those who control it. A novel approach to enhanced privacy-preserving techniques, brought to market using federated learning, addresses these concerns directly. Furthermore, as computing shifts to models running at the edge, the orchestration of cloud and edge becomes critical to long-term energy requirements. A decay in performance improvement from classical chip scaling and an increasing share of energy consumption from AI data centers heighten concerns about power consumption over time.

Budget limitations, infrastructure availability, bandwidth constraints, and real-time requirements drive much of the intelligence toward the edge of the network—that is, the vicinity of where the data is generated. These edge devices sense, act, and infer based on learned models that meet the energy budgets, static latency thresholds, and increasing workload support requests. Productive orchestration of the resulting distributed computational resources uses federated learning, where aggregates of model updates are collected and combined rather than the raw training data, often enabling a degree of differential privacy.

Classical data transfers are reduced, improving operational efficiency and reducing costs by constraining burdens on the core and regional fabric of the cloud. Such constraints also create opportunities to offer confidentiality to sensitive data, especially in regulated regimes. Nevertheless, privacy-preserving operational models add complexity to model deployment, hardware infrastructure needs to support privacy-preserving inference using the outputs of the federated learning process, and privacy-preserving inference models often require elevated processing capability. Availability of hardware accelerators at the predicted required price points and levels of performance will therefore strongly influence the commercial adoption of these models.

Final Thought

Learn more with Minterminds.