Analog Devices,inc.

Mitsubishi Electric Research Laboratories, Inc.

Other

Benjamin Vigoda

Jeffrey G Bernstein

David C Reynolds

Alexander Alexeyev

William Bradley

Eric G Nestler

Vladimir Zlatkovic

Jeffrey Venuti

David MacDonald Hossack

Shawn Hershey

Robert W Adams

Martin McCormick

Kartik Nanda

Andrew Schweitzer

Mira Wilczek

Theophane Weber

Timothy J Caputo

William Peet

Rishi Chaturvedi

Joshua Park

An inference task is performed using a computation device having a plurality of processing elements operable in parallel and connected via a connectivity system. Performing the task includes accepting at the device a specification of at least part of the inference task. The specification characterizes a plurality of variables and a plurality of factors, each factor being associated with a subset of the variables. Each of the processing elements is configured with data defining one or more of the plurality of factors. At each of the processing elements, computation associated with one of the factors is performed concurrently with other of the processing elements performing computation associated with different ones of the factors. Messages are exchanged via a connectivity system. The messages provide inputs and/or outputs to the processing elements for the computations associated with the factors and provide a result of performing of the at least the part of the inference task.

Programmable probability processing

In an aspect, in general, a programmable computation device performs computations of an inference task specified by a plurality of variables and a plurality of factors, each factor being associated with a subset of the variables. The device includes one or more processing elements. Each processing element includes a first storage for a definition of a factor, a second storage for data associated with the inputs and/or outputs of at least some of the computations, and one or more computation units coupled to the first storage and the second storage for performing a succession of parts of the at least some of the computations that are associated with a factor, the succession of parts defined by data in the storage for the definition of the factor.

Microphone stages in a microphone array may be coupled together in a daisy chain. Each stage may include a microphone, an analog to digital converter, a decimation unit, a receiver, an adder, and a transmitter. The converter may convert analog audio microphone signals into digital codes that may be decimated. The adder may add decimated digital codes in each stage to a cumulative sum of decimated digital codes from prior stages. This new sum may be transmitted to the next microphone stage, where the adder may add the decimated digital codes from that stage to the cumulative sum. A serial interface may be used to connect the transmitters and receivers of each of the stages. The serial interface may be used to transmit the cumulative sum of decimated digital codes between the stages. The serial interface may also be used to transmit configuration data between the stages.

Microphone array with daisy-chain summation

In a data processing system, a method for implementing a factor graph having variable nodes and function nodes connected to each other by edges includes implementing a first function node and a on a first computer system, the first computer system being in network communication with a second computer system; establishing a network connection to each of a plurality of processing systems; receiving, at the first function node, soft data from a variable node implemented on one of the processing systems, the soft data including an estimate of a value and information representative of an extent to which the estimate is believed to correspond to a correct value; and transmitting, from the first function node to the one of the processing systems, soft data representing an updated estimate of the value.

Distributed factor graph system

A circuit implementing a soft logical processing network includes an interconnection of analog processing elements, which can include soft logic gates, for instance soft Equals gates and soft XOR gates. In some examples, each of the soft logic gates include multiple circuit parts, with each part including an input configured to accept a voltage signal representation of a soft logical quantity, and a conversion section configured to use the accepted voltage representation to form a corresponding current signal. The current signals are combined to form a signal representation of the output of the gate. In an application of soft logic gates, a memory includes a group of electrical storage elements, each electrical storage elements carrying a respective storage values; a group of conversion elements, each conversion element being coupled to a respective electrical storage element for selectively converting the corresponding storage value to a current signal; and a current combination element for combining the current signals to form an output signal.

Circuits for soft logical functions

A storage device includes a storage array having a group of storage elements. Each storage element can written to a discrete set of physical states. A read circuit selects one or more storage elements and generates, for each selected storage element, an analog signal representative of the physical state of the selected storage element. A signal processing circuit processes the analog signal to generate a plurality of outputs, with each output representing a degree of an association of the selected storage element with a different subset of one or more of the discrete set of physical states.

Storage devices with soft processing

A processor includes a first memory module for storing a first set of storage values each representing a respective input, and a second memory module for storing a second set of storage values in analog form. An analog module is coupled to the first and the second memory modules. The analog module is configured to, in each operation cycle of at least one iteration, update at least some of the second set of storage values based on the first and the second sets of storage values. An output module is for generating a set of outputs from at least some of the second set of storage values.

Belief propagation processor

A processor implements a network of functional nodes and communication paths between the nodes. The processor includes a plurality of circuit implementations of the functional nodes of the processor; and a plurality of signal paths implementing the communication paths linking the circuit implementations of the nodes. At least some of the signal paths are configured to pass signal values represented according to temporal patterns of signal levels on the signal paths. The processor also includes a plurality of circuit components for conversion between a signal value represented as a signal level (e.g., voltage or current level) and a signal value represented as a temporal pattern.

Mixed signal stochastic belief propagation

In one aspect, a storage device includes a plurality of storage strings, each comprising a serial interconnection of a plurality of active storage elements, each storage element having a part for maintaining a storage state and a part of modulating a current through the element according to the storage state. The device also includes mapping circuitry for selectively sensing a storage state of a storage element in a storage string by forming current though the storage element that is a non-linear function of the storage state. In some examples, the mapping circuitry comprises reference string of active elements, and the mapping circuitry selectively senses a storage state by forming a difference in currents in the sensed storage string and in the reference string that is a non-linear function of the storage state. In some examples, the active storage elements comprise floating gate transistors.

Storage device architecture

An approach to signal conversion adapts the signal conversion process, for example, by adapting or configuring signal conversion circuitry, according to inferred characteristics (e.g., probability distribution of value) of a signal being converted. As an example, an analog-to-digital converter (ADC) may be adapted so that its accuracy varies across the range of possible input signal values in such a way that on average the digital signal provides a higher accuracy than had the accuracy remained fixed. In another example, models (and corresponding inference circuitry) of both an input signal process and of a quantization process are used to improve signal conversion accuracy.

Signal conversion

The process of implementing a belief propagation network in software and/or hardware can begin with a factor-graph-designer who designs a factor graph that implements that network. A development system provides a user with a way to specify a factor graph at a high or abstract level, and then solve the factor graph, or make an instance of the factor graph in software and/or hardware based on the specification. Factor graphs enable designers to create a graphical model of complicated belief propagation networks such as Markov chains, hidden Markov models, and Bayesian networks.

Implementation of factor graphs

A circuit includes a signal processing circuit for accepting an input and for generating a set of outputs. The input is provided in an input range that has a set of representative values, and each output represents a measure of an association of the input with one or more of the representative values. The signal processing circuit includes a group of output sections, each output section being responsive to the input of the signal processing circuit. Each output section includes one or more sigmoid generators. Each sigmoid generator is responsive to an input of the output section to generate an output that represents a sigmoid function of the input of the output section. Each output section also includes a circuitry for combining the outputs of the one or more sigmoid generators to form one of the set of outputs of the signal processing circuit. An input transformation circuit is coupled to the plurality of output sections. The input transformation circuit is configurable to transform the input of the signal processing circuit for controlling a mapping characteristic from the input to the set of outputs.

Signal mapping

Some general aspects relate to systems and methods of analog computation using numerical representation with uncertainty. For example, a specification of a group of variables is accepted, with each variable having a set of at least N possible values. The group of variables satisfies a set of one or more constraints, and each variable is specified as a decomposition into a group of constituents, with each constituent having a set of M (e.g., M&#x3c;N) possible constituent values that can be determined based on the variable values. The method also includes forming a specification for configuring a computing device that implements a network representation of the constraints based on the specification of the group of variables. The network representation includes a first set of nodes corresponding to the groups of constituents, a second set of nodes corresponding to the set of constraints, and interconnections between the first and the second sets of nodes for passing continuous-valued data.

Growing community of inventors

Winchester, MA, United States of America

Benjamin Vigoda