Benefits of TinyML
TinyML enables the deployment and on-device processing of ML models into resource constrained edge devices. TinyML entails numerous benefits that are discussed below.
Data collected by edge device sensors is processed by an on-device ML model to generate the results. Tiny ML models are optimized to have a reduced number of parameters and there is no time wasted in transmitting the data to a datacenter server or cloud for processing. This in turn results in a very low latency (i.e., fast turnaround time) from data collection to result generation. TinyML applications can work in real time due to their low latency for mission critical applications.
Machine learning consumes data. Privacy has been a major concern against the adoption of machine learning technologies since consumers are oftentimes unwilling to share their data or transport it over the internet due to security concerns. TinyML enables on-device processing of data in real time. In TinyML applications, since data is neither transmitted out of device nor stored anywhere, there are no privacy concerns.
Fingerprints used for biometric authentication is an example of data with privacy value. Fingerprints are typically represented as 129 points. If this fingerprint data is compromised over the internet, it will be an irreparable loss. Passwords can be changed, fingers can’t be! That’s the reason sensitive biometric data must always be stored and processed on-device.
TinyML applications are expected to run on resource constrained devices, process real time data and produce results within data sampling intervals. Driven by all these requirements TinyML models are optimized to run with fewer parameters, use reduced computations and consume low power.
TinyML applications mostly use on-device sensors’ data. Majority of the sensor data (other than microphone audio data and camera video data) volumes are small (less than 1K bytes/sec). Consequently, machine learning models to process these data are also light.
TinyML applications cater to small, resource constrained devices with poor to no internet connectivity. On-device sensors capture the data, data is processed on-device and hence there is no raw sensor data transmission bandwidth involved. Sometimes Inference or analytics results limited to 100s of bytes are transmitted to cloud requiring very low bandwidth.
On-device execution of TinyML models makes it very appealing connectivity denied and degraded environments. They are perfect match to stealth mode operation.