Joshua Bloom

While machine-learning based inference on edge devices is becoming more tractable with falling costs of compute and storage, there are security and privacy considerations where a hub-spoke approach is preferable. For high-value use cases in particular, the centralization of a machine-learned model (e.g., in a cloud environment) allows for a concerted defensive posture on a smaller footprint than if a model is distributed to all nodes; at the edge, exposure to heterogenous network environments… Show more

While machine-learning based inference on edge devices is becoming more tractable with falling costs of compute and storage, there are security and privacy considerations where a hub-spoke approach is preferable. For high-value use cases in particular, the centralization of a machine-learned model (e.g., in a cloud environment) allows for a concerted defensive posture on a smaller footprint than if a model is distributed to all nodes; at the edge, exposure to heterogenous network environments make security of that model more intractable. The downside to centralization however is the need to transmit raw data over networks, making privacy of those data (following from the loss of governance once data leaves a site) a major concern. Here we demonstrate an approach that simultaneously addresses privacy and security to compress and obscure data in an edge-to-cloud transmission setting. We use autoencoding deep neural networks to train unsupervised models on domain-specific datasets (in this case, large corpora of medical images). The encoding portion of trained networks are transmitted to edge devices and the encoding of new data are transmitted to another device that contains the decoding portion of the network. The original data can be reconstructed with a tunable loss and used for prediction or to produce new models. An attack on the edge would not yield enough information to reconstruct the full output of the model; likewise an attack on the transmission (even if standard encryption is subverted) would not yield enough information to reconstruct the original data. Show less

The spectrometer of the present invention comprises a series of layers consisting of alternating pixilated detection and transmission optics, with a collimating array of lenslets at top. Radiation projected through the collimating arrays of lens passes through the transmission optics where, at interfaces acting as dichroics at certain designated wavelengths, are reflected towards the detection arrays. Light redward of the reflection at each interface is transferred farther through the… Show more

The spectrometer of the present invention comprises a series of layers consisting of alternating pixilated detection and transmission optics, with a collimating array of lenslets at top. Radiation projected through the collimating arrays of lens passes through the transmission optics where, at interfaces acting as dichroics at certain designated wavelengths, are reflected towards the detection arrays. Light redward of the reflection at each interface is transferred farther through the transmission optics. The intensity of the incident light at each pixel in each detection layer is recorded, providing a spatially and spectrally resolved representation of the incoming radiation. The spectral resolution is set by the number of dichroic interfaces. The transmissive optics, if made of a solid material, may act also as a solid-body thermal transport for potential cooling of the detection layers. It allows for spatially resolved spectroscopy from a compact device that has a form factor not significantly different than traditional imagers. In other implementations, spectral information of a similar imaging resolution is obtained at significantly reduced efficiency and may result in the spectral information being obtained non-simultaneously (this is disadvantageous when the goal is to the study the spatio-temporal dynamics of the system being images). This device allows the simultaneous capture across the visible portion of the electromagnetic bandpass, allowing for spatio-temporal-spectral dynamics to be captured. Show less

Described is an apparatus for early warning of natural and manmade
disasters. Preferably, the apparatus is connected through the Internet to a
provider of disaster warning information. A user of the apparatus determines the
geolocation where the apparatus will be operated and subscribes to one or more
alerting channels for that geolocation. Preferably the subscription is established
and stored via a cloud-based interface. Depending on the nature of the disaster
and the… Show more

Described is an apparatus for early warning of natural and manmade
disasters. Preferably, the apparatus is connected through the Internet to a
provider of disaster warning information. A user of the apparatus determines the
geolocation where the apparatus will be operated and subscribes to one or more
alerting channels for that geolocation. Preferably the subscription is established
and stored via a cloud-based interface. Depending on the nature of the disaster
and the distance from the origin of the disaster, alerts may be generated seconds
(earthquakes), minutes (tornados), hours (tsunamis) before the effects are
noticed local to the device. The user can set thresholds for alert severity, and,
via audible and/or visual alerts for example, the apparatus can warn the user of
the expected arrival of local shaking due to earthquakes and/or other deleterious
effects due to other natural/manmade disasters and events. The technology
leverages a nascent, but growing set of sensor networks that discover/predict
and broadcast (including via standard internet protocols) such events.
In various embodiments, the apparatus may be mounted in a home,
classroom, or office just as would a smoke alarm. The electronics in the
apparatus may be powered from a low-voltage, low-power supply (e.g., 5 volt or
3 volt input) such as provided by commonly available USB interfaces. Near-continuous
operation of the apparatus may be maintained through an internal
battery (with pass-through charging) that allows the apparatus to run for days
after external power is lost.
Preferably the apparatus maintains a constant connection to at least one
network (such as a home wireless LAN), which is in turn connected to remote
servers through the Internet. The apparatus may also include the capability to
establish/maintain a new connection to a new network using WiFi or cellular or
radio technologies to assure robust alerting. Show less