: Applying machine learning in embedded systems. Since Hamming distance is very compact, and neighbours can be found fast. Specially HDML / Hamming Distance Metric Learning (Norouzi 2012), Convolutional Neural networks (quantized). Should both be useful for typical tasks and efficiently implementable. 
In audio-processing, could we use a speech detection algorithm to avoid storing samples with speech in them?Ĭan then store/transmit the other data in order to do quality assurance and/or further data analysis. While feature hashing is ideally suited to categorical features, it also empirically works well on continuous features The typical motivation for using the technique is a reduction in memory requirements or the ability to perform stateless feature extraction. Hashing trick is an established way of processing data as part of training a machine learning model. Using feature hashing on client/sensor-side, before sending to server that performs training. Scalable Machine Learning with Fully Anonymized Data. TODO: cost (monetary) of data transmission, for different wireless techsĭoing more of the data processing locally, enables storing or transmitting privacy sensitive data more seldom. TODO: overview of sending range, for different wireless tech TODO: overview of data transmission capacity, for different wireless tech TODO: overview of typical energy requirements, for different wireless tech Never connected to charger, should run forever Periodically used on battery, else plugged in. Material identification using reflecive spectrometer 1. Electronic nose using arrays of MEMS detectors. Adaptive signalling and routing for wireless transmission in Wireless Sensor networks. Environmental monitoring, using microphone to detect unwanted activity like cutting down trees. Monitoring eating activity using accelerometer 1. Health status of animals via activity detected using accelerometer. Normally sending day/week aggregates, on event/anomaly detection send data immediately Speech/command recognition as human input device, using microphone. 
Gesture recognition as human input device, using accelerometer/gyro data.Anomaly/change detection for predictive maintenance, using audio/vibration data, or electrical data.Appliance disaggregation, using aggregated power consumption data.Activitity detection for people, using audio/accelerometer data.
Predictive maintenance, using audio/vibration data. Full/raw sensor data is not valuable to store. Sending raw sensor data has privacy implications.Īudio, video, accelerometer/IMU, current sensor, radiowaves. If gateways are used, they mostly forward communication (no data processing). The defaults right now are to do conventional signal processing (no learning) in sensor,Īnd stream raw data to the cloud for storage and processing. Background What and when to use machine learning nnom - Fixed-point neural network compiler for microcontrollers. Lower time to market, enable more developers Best practices underdocumented (or underdeveloped?). Very little documentation of entire development process.įrom planning, data aquisition, model design. "Small DNN" work mostly on computer vision for mobile phones (model size 1000x of uC). Non-neural models missing inference engines designed for microcontrollers. Neural models lacking for non-ARM micros. Lots of research and many announcements of low-power co-processors, but little on market yet. Human activity detecton on accelerometers. 
Keyword-spotting/wake-word on audio well established.Basic tools available for converting Tensorflow models.A few special-designed ML algorithms made.Some implementations available for non-neural models that can be used.Ex: CNN and RNN in CMSIS-NN, FC in uTensor Deep models have efficient implementations for ARM Cortex-M.Of ML inference on general-purpose microcontrollers. Training phase can run on a standard computer/server, using existing tools as much as possible. Focused primarily on running inference/prediction/feed-forward part on a microcontroller (or small embedded device).
0 kommentar(er)
