Artificial neural network processing methods and system

A device, comprising:

• a memory;
• and
• processing circuitry coupled to the memory, wherein the processing circuitry, in operation:
• sets a dimensional parameter (d) of a lattice to an integer value, the lattice having a plurality of lattice points and identified by a set of basis vectors (b1, b2);
• selects a set of weight parameters (wi) of a respective processing layer of a plurality of processing layers of an artificial neural network (ANN), the ANN having a plurality of processing layers with respective processing layer parameters, the processing layer parameters including at least one set of weight parameters (wi), at least one input activation parameter (aj), at least one output activation parameter (bj) and at least one activation function parameter (ρ);
• applies vectorization processing to the selected set of weight parameters (wi), producing a set of weight vectors (u1, u2, u3), and arranging the set of weight vectors (u1, u2, u3) as a matrix of weight vectors (U);
• performs normalization processing of the matrix of weight vectors (U), producing a matrix of normalized weight vectors (U′);
• applies lattice vector quantization (LVQ) processing to the matrix of normalized weight vectors (U′), producing a codebook of codewords;
• and
• indexes the produced codebook, the indexing comprising encoding codewords of the codebook as a function of the lattice, producing respective tuples of indices (is, ia, iaLUT), wherein performing the normalization processing of the matrix of weight vectors (U) comprises solving an optimization problem having:
• a first term to provide normalized weight values which approximate the at least one output activation parameter (bj) of the ANN as a function of the at least one input activation parameter (aj) of the ANN;
• and
• a regularization term ΩL(u) to amplify normalized weight values distanced from the lattice points of the selected lattice.

A system, comprising:

• a memory, which, in operation, stores compressed normalized weight values;
• a processing core;
• and
• a hardware accelerator engine coupled to the memory and the processing core, wherein the hardware accelerator engine, in operation:
• decompresses stored normalized weight values, producing decompressed weight values;
• and
• performs an operation of an artificial neural network (ANN) as a function of said decompressed weight values, wherein the stored compressed normalized weight values correspond to a solution to an optimization problem having:
• a first term to provide normalized weight values which approximate at least one output activation parameter (bj) of the ANN as a function of at least one input activation parameter (aj) of the ANN;
• and
• a regularization term ΩL(u) to amplify normalized weight values distanced from lattice points of a selected lattice.

A system, comprising:

• a memory, which, in operation, stores compressed normalized weight values;
• and
• processing circuitry coupled to the memory, wherein the processing circuitry, in operation:
• implements an artificial neural network;
• decompresses stored normalized weight values, producing decompressed weight values;
• and
• performs an operation of the ANN as a function of the decompressed weight values, wherein the stored compressed normalized weight values correspond to a solution to an optimization problem having:
• a first term to provide normalized weight values which approximate at least one output activation parameter of the ANN as a function of at least one input activation parameter of the ANN;
• and
• a regularization term to amplify normalized weight values distanced from lattice points of a selected lattice;
• and
• classifies an image using the ANN, the classifying including performing the operation of the ANN.