Confidence Preservation Property in Knowledge Distillation Abstractions

Social media platforms prevent malicious activities by detecting harmful content of posts and comments. To that end, they employ large-scale deep neural network language models for sentiment analysis and content understanding. Some models, like BERT, are complex, and have numerous parameters, which makes them expensive to operate and maintain. To overcome these deficiencies, industry experts employ a knowledge distillation compression technique, where a distilled model is trained to reproduce the classification behavior of the original model. The distillation processes terminates when the distillation loss function reaches the stopping criteria. This function is mainly designed to ensure that the original and the distilled models exhibit alike classification behaviors. However, besides classification accuracy, there are additional properties of the original model that the distilled model should preserve to be considered as an appropriate abstraction. In this work, we explore whether distilled TinyBERT models preserve confidence values of the original BERT models, and investigate how this confidence preservation property could guide tuning hyperparameters of the distillation process.

Highly curved image sensors: a practical approach for improved optical performance

The significant optical and size benefits of using a curved focal surface for imaging systems have been well studied yet never brought to market for lack of a high-quality, mass-producible, curved image sensor. In this work we demonstrate that commercial silicon CMOS image sensors can be thinned and formed into accurate, highly curved optical surfaces with undiminished functionality. Our key development is a pneumatic forming process that avoids rigid mechanical constraints and suppresses wrinkling instabilities. A combination of forming-mold design, pressure membrane elastic properties, and controlled friction forces enables us to gradually contact the die at the corners and smoothly press the sensor into a spherical shape. Allowing the die to slide into the concave target shape enables a threefold increase in the spherical curvature over prior approaches having mechanical constraints that resist deformation, and create a high-stress, stretch-dominated state. Our process creates a bridge between the high precision and low-cost but planar CMOS process, and ideal non-planar component shapes such as spherical imagers for improved optical systems. We demonstrate these curved sensors in prototype cameras with custom lenses, measuring exceptional resolution of 3220 line-widths per picture height at an aperture of f/1.2 and nearly 100% relative illumination across the field. Though we use a 1/2.3" format image sensor in this report, we also show this process is generally compatible with many state of the art imaging sensor formats. By example, we report photogrammetry test data for an APS-C sized silicon die formed to a 30$^\circ$ subtended spherical angle. These gains in sharpness and relative illumination enable a new generation of ultra-high performance, manufacturable, digital imaging systems for scientific, industrial, and artistic use.