Massachusetts Institute of Technology

Hewlett-packard Development Company, L.P.

Lucent Technologies Inc.

International Business Machines Corporation

Toyota Research Institute, Inc.

Picturetel Corporation

Gregory W Wornell

Brian Chen

John George Apostolopoulos

Mitchell D Trott

William T Freeman

Guy Rosman

Daniela Rus

Sertac Karaman

Emin Martinian

Alan V Oppenheim

Fredo Durand

Igor Gilitschenski

Ying-zong Huang

Uri Erez

Ram Zamir

Antonio Torralba

Alexander Amini

Prasanna Sattigeri

Albert Chan

Felix Maximilian Naser

Henrique S Malvar

Joshua Ka-Wing Lee

Gary J Sullivan

Examples described herein provide a computer-implemented method that includes training, by one or more processing devices, a first neural network for classification based on training data in accordance with a first learning objective, the first neural network producing an intermediate feature function and a final feature function as outputs. The computer-implemented method further includes training, by the one or more processing devices, a second neural network for classification based on the intermediate feature function and the final feature function and further based at least in part on target task samples in accordance with a second learning objective. Training the second neural network includes computing maximal correlation functions of each of the intermediate feature function, the final feature function, and the target task samples.

Multi-source transfer learning from pre-trained networks

The present disclosure provides systems and methods to detect occluded objects using shadow information to anticipate moving obstacles that are occluded behind a corner or other obstacle. The system may perform a dynamic threshold analysis on enhanced images allowing the detection of even weakly visible shadows. The system may classify an image sequence as either 'dynamic' or 'static', enabling an autonomous vehicle, or other moving platform, to react and respond to a moving, yet occluded object by slowing down or stopping.

Systems and methods of detecting moving obstacles

An encoder generates a compressed data sequence from an original data sequence using many-to-one mapping independently of a source model associated with the original data sequence and without extracting the source model. A decoder uses both the source model associated with the original data sequence and the mapping applied during compression that is devoid of, in substance, the source model, to regenerate, at least in part, the original uncompressed data sequence from the compressed data sequence that does not include a significant portion of the source model.

Systems and methods for model-free compression and model-based decompression

A powerful new class of methods for encoding digital data for reliable transmission over unreliable communication channels is described. With this method, the message bits are divided into multiple submessages and the bits in each layer are encoded using a standard error correction code to provide a plurality of subcodewords. A first linear transformation is applied to each of the subcodewords. The so-transformed subcodewords from the different submessages are then combined to form a first redundancy block to be transmitted. Additional redundancy blocks are generated by repeating this process on the same message but with jointly related nonidentical sets of linear transformations. The result is a set of codewords for each message which are then used to generate a transmitted waveform in one of several different ways, depending upon the application. For example, the codewords can be used in sequence (temporally) to form a rateless code for transmission over channels of unknown quality, in which case a receiver accumulates as many of the codewords as are necessary to successfully decode. As another example, the codewords can be spatially distributed if the channel has multiple-inputs, as would be the case in a multi-antenna system. As still another example, the codewords can be distributed over subchannels in a frequency-division multiplexing system.

System and apparatus for error control codes based on layering and linear transformations

Described are techniques for performing lossy encoding. Source data and quality data are received by an encoder. The encoder maps the source data into a compressed representation having a level of distortion in accordance with the quality information. The compressed representation may be decoded without using the quality information.

Lossy data compression exploiting distortion side information

A block-iterative equalizer is adapted for use in contemporary digital communication system receivers. In a preferred embodiment, data received over a communication channel is processed by a linear feed-forward filter and the resulting filtered signal is provided to a slicer which makes a first set of tentative symbol decisions. During later iterations, the same received data is processed by the linear feed-forward filter, the feed-forward filter parameters being modified at each iteration based on the received data and the tentative decisions are themselves filtered by a second, 'feed-back' filter and used to improve the tentative decisions.

Block-iterative equalizers for digital communication system

Communication over lossy packet networks such as the Internet is hampered by limited bandwidth and packet loss. The present invention provides a path diversity transmission system for improving the quality of communication over a lossy packet network. The path diversity transmission system explicitly sends different subsets of packets over different paths, thereby enabling the end-to-end application to effectively see an average path behavior. Generally, seeing this average path behavior provides better performance than seeing the behavior of any individual random path. For example, the probability that all of the multiple paths are simultaneously congested is much less than the probability that a single path is congested. The resulting path diversity can provide a number of benefits, including enabling real-time multimedia communication and simplifying system design (e.g., error correction system design). Two exemplary architectures for achieving path diversity are described herein. The first architecture is based on source routing, and the second architecture is based on a relay infrastructure. The second architecture routes traffic through semi-intelligent nodes at strategic locations in the Internet, thereby providing a service of improved reliability while leveraging the infrastructure of the Internet.

Method and system for packet communication employing path diversity

A compression and transmission method and system for use to communicate information from a first device to a second device in a network of devices. First, the first device receives information to be communicated to the second device. Second, the first device compresses the information to be communicated by employing prior information of the first device and prior information of the second device. The compressed information is then transmitted to the second device.

Distributed compression and transmission method and system

A system, method, and product provide distortion compensation when watermarking a host signal with a watemark signal. The watermark signal includes watermark-signal components having watermark-signal values, and the host signal includes host-signal components having host-signal values. A system includes an ensemble designator that designates a plurality of embedding generators, each corresponding to a single watermark-signal value of a co-processed group of one or more watermark-signal components. Also included is an embedding value generator that generates embedding values by each embedding generator. The system also has a point coder that sets at least one host-signal value of one or more selected host-signal components to a first embedding value to form a composite-signal value. A distortion compensator modifies at least one composite signal value. The distortion compensator includes a difference signal generator that generates a difference signal component having a first difference-signal value that is the difference between a host-signal value and its corresponding composite-signal value. The distortion compensator also includes a difference signal processor that scales the difference-signal value by a scaling factor, A distortion compensation combiner then generates a distortion-compensated composite signal based on the scaled difference-signal value and the composite-signal value. The distortion compensator may modify the composite signal value based on the number of possible watermark signal values to be embedded, a watermark-induced distortion level, an expected channel-induced distortion level, and a reliability of extracting the watermark signal.

System, method, and product for distortion-compensated information embedding using an ensemble of non-intersecting embedding generators

A system, method, and product are provided to (1) embed a watermark signal into a host signal, thereby generating a composite signal, (2) optionally enable the composite signal to be transmitted over a communication channel, and (3) optionally extract the watermark signal from the transmitted composite signal. In one embodiment, the invention is a method for watermarking a host signal with a watermark signal. The watermark signal is made up of watermark-signal components, each having one of two or more watermark-signal values. The host signal is made up of host-signal components, each having one of two or more host-signal values. The method includes: (1) generating two or more embedding generators, each corresponding to a single watermark-signal value of a co-processed group of one or more watermark-signal components; (2) having each embedding generator generate two or more embedding values, the total of which is referred to as an original embedding-value set such that at least one embedding value generated by one embedding generator is different than any embedding value generated by another embedding generator; and (3) setting a host-signal value of one or more selected host-signal components to an embedding value of a particular embedding generator, thereby forming a composite-signal value, such that the particular embedding generator corresponds to the watermark-signal value of the co-processed group of watermark-signal components, and such that the embedding value of the particular embedding generator is selected based on its proximity to the host-signal value.

System, method, and product for information embedding using an ensemble of non-intersecting embedding generators

A system, method, and product are provided to (1) pre-process one or more primary signals to generate a transformed host-signal and/or a transformed watermark-signal; (2) embed one or more watermarked signals and/or transformed watermark signals into a host signal and/or the transformed host signal, thereby generating a composite signal, (2) optionally enable the composite signal to be transmitted over a communication channel, and (3) optionally extract the watermark signal from the transmitted composite signal. An embedding value may be the closest of all embedding values generated by an embedding generator to a host-signal value that is to be quantized. Embedding values may be based on a trellis-coded pre-determined relationship between embedding values, or on predetermined relationships based on lattice quantization. The method may also include a fourth step of extracting the first watermark-signal value from a composite-signal value to form a reconstructed watermark-signal value. The present invention may also implement adaptive embedding and, in some implementations, super-rate quantization. For example, the invention may be a system that includes an ensemble designator that designates a plurality of adaptive embedding generators, each corresponding to a single watermark-signal value of a co-processed group of one or more watermark-signal components. Also included in this implementation is an adaptive embedding value generator that generates, by each adaptive embedding generator, a plurality of adaptive embedding values.

System method, and product for information embedding using an ensemble of non-intersecting embedding generators

A method and apparatus for encoding a sampled signal using a nonscalar vector quantizer. A lapping window operator is applied to blocks of signal samples of length greater than N to produce length-N transform vectors which are then encoded by a nonscalar vector quantizer. The method and apparatus may use or have a window operator that is identical to that of the modulated lapped transform, that is a 2N.times.N matrix operator, that is identical to that of the modulated lapped transform with asymmetrical, orthogonal butterflies, or that has butterfly transmittances given by h�n!=sin�(2n+1).pi./4N!. The method and apparatus may use or have a vector quantizer is an unconstrained mean-square optimized vector quantizer having codewords of length N. The method and apparatus may use or have a vector reconstructor being an inverse of the nonscalar vector quantizer to reconstruct length-N vectors coupled through a transmission channel to received the length-N vectors from the vector quantizer, and a block decoder connected to receive as input the reconstructed length-N vectors from the vector reconstructor and to apply a reconstructor window operator to produce vectors that may be superimposed to generate a reconstruction of the sampled signal. The reconstructor window operator may be the inverse of the lapping window operator.

Lapped orthogonal vector quantization

Method and apparatus for preparing a message stream of symbols for wireless transmission by pre-coding it through convolution with a characteristic 'signature' of predetermined symbols identifying a particular user of the message system. The time duration of the signature should exceed the inter-symbol interval of the message stream in order to reduce the likelihood of error caused by fading in the transmission channel.

Spread-response precoding system having signature sequences longer than

Transmitting and receiving method and apparatus especially suited to multi-user communication systems such as cellular systems, in which multi-path transmission encounters time-variable fading characteristics in the channel. Pre-coding message stream of symbols by convolution with predetermined signature sequences for respective users enables separation of respective user messages and combats effects of variable fading.

Growing community of inventors

Wellesley, MA, United States of America

Gregory W Wornell