Machine trained network using novel coding techniques

A non-transitory machine readable medium storing a program for analyzing an input image to determine whether the input image contains a picture of an object belonging to one object category from a plurality of object categories, the program comprising sets of instructions for:

• propagating an input image through a plurality of layers of a machine-trained (MT) network to produce a multi-dimensional codeword representing the input image in a multi-dimensional space that includes a first set of codewords comprising at least one codeword for each object category in the plurality of object categories, wherein each codeword in the first set of codewords represents one of the object categories, wherein propagating the input image through the plurality of layers comprises providing the input image to a first input layer of processing nodes that generate a first set of intermediate outputs that are subsequently provided to at least one subsequent layer of the MT network, each subsequent layer of the MT network comprising a respective set of processing nodes that receive intermediate outputs from at least one previous layer of the MT network and generate a respective set of intermediate outputs until an output layer of the MT network generates the multi-dimensional codeword representing the input image;
• for the produced codeword representing the input image, generating a signature comprising a set of affinity scores with each affinity score identifying a proximity of the produced codeword to a different codeword in a second set of previously defined codewords representing other images, wherein each codeword in the second set of codewords is located at least a particular distance in the multi-dimensional space from any of the codewords in the first set of codewords in order to define boundaries between the codewords in the first set of codewords representing the object categories in the multi-dimensional space;
• comparing the signature generated for the produced codeword representing the input image with previously defined signatures generated for the first set of codewords by computing, for each particular first-set codeword, a set of affinity scores that comprises a plurality of affinity scores each of which identifies a proximity of the particular first-set codeword representing an object category to a different codeword in the second set of codewords defining boundaries between the object categories;
• and
• based on said comparison, identifying an object category to associate with the input image, said association specifying that the input image contains a picture of an object that belongs to the identified object category.

A machine-trained network for analyzing an input image to determine whether the input image contains a picture of an object belonging to one object category from a plurality of object categories, the network comprising:

• a plurality of layers of machine-trained processing nodes through which the input image is propagated to produce a multi-dimensional codeword representing the input image in a multi-dimensional space that includes a first set of codewords comprising at least one codeword for each object category in the plurality of object categories, wherein each codeword in the first set of codewords represents one of the object categories, wherein the input image is propagated through the plurality of layers by a first input layer of processing nodes receiving the input image and generating a first set of intermediate outputs that are subsequently provided to at least one subsequent layer of processing nodes, each subsequent layer of processing nodes comprising a respective set of processing nodes that receive intermediate outputs from at least one previous layer of processing nodes and generate a respective set of intermediate outputs until an output layer of processing nodes generates the multi-dimensional codeword representing the input image;
• a scoring engine to generate a set of affinity scores for the produced codeword representing the input image, each affinity score in the affinity score set identifying a proximity of the produced codeword to a different codeword in a second, different set of previously defined codewords representing other images, wherein each codeword in the second set of codewords is located at least a particular distance in the multi-dimensional space from any of the codewords in the first set of codewords in order to define boundaries between the codewords in the first set of codewords representing the object categories in the multi-dimensional space;
• and
• a comparator (i) to compare the set of affinity scores generated for the produced codeword representing the input image with sets of affinity scores previously generated for the first-set codewords, wherein each set of affinity scores that is previously generated for each particular first-set codeword comprises a plurality of affinity scores each of which identifies a proximity of the particular first-set codeword representing an object category to a different codeword in the second set of codewords defining boundaries between the object categories, and (ii) based on the comparison, to identify an object category to associate with the input image, said association specifying that the input image contains a picture of an object that belongs to the identified object category.