UBER and Intel’s Machine Learning platforms

Optimization

Cormac Brick from Intel discusses advanced optimization techniques for neural networks, emphasizing their importance for diverse silicon architectures. He highlights four key methods: model pruning, quantization, sparsification, and weight compression, which enhance performance across various hardware, including VPUs, GPUs, and FPGAs 1. These techniques ensure models remain versatile and efficient, regardless of the silicon used, and are crucial for maximizing performance on specific devices 1.

If you employ these techniques, you're not going to hurt your model's ability to run across a broad range of silicon.

--- Cormac Brick

Brick also mentions Intel's Distiller project, which consolidates fragmented optimization tools into a more accessible format, aiding developers in applying these techniques effectively 2.

   

Domain Transfer

Domain transfer in neural networks is a powerful approach for tailoring models to specific applications, as explained by Cormac Brick. He notes that many existing models are optimized for complex datasets like ImageNet, which may not align with real-world needs 3. By employing techniques like channel pruning and sparsification, developers can simplify models for tasks such as recognizing common household objects, enhancing efficiency and performance on edge devices 3.

You can get away with a much simpler network with a less representational capacity to solve that 100 image problem.

--- Cormac Brick

These methods allow for more efficient use of resources, particularly in bandwidth-limited environments, making them ideal for embedded systems 3.