A method and a system for classifying structured data

Abstract

The present subject matter discloses a method (900) for classifying data. The method (900) includes receiving a plurality of feature vectors associated with data as an input. The method (900) includes performing an adaptive stream partitioning of the plurality of features into a plurality of streams. The method (900) includes processing the plurality of streams parallelly to one another to generate a plurality of processed streams. The method (900) includes performing a sparse adaptive fusion on the plurality of processed streams to concatenate the plurality of processed streams into a fused vector. The method (900) includes classifying the data based into one or more classes based on a probability assigned to each stream in the fused vector (FIG. 3)

Description

TECHNICAL FIELD[1] The present subject matter relates to a field of data classification, particularly, thepresent subject relates to a method for classifying structured data in a neural networkarchitecture.BACKGROUND OF THE INVENTION[2] The background description includes information that may be useful in understandingthe present subject matter.[3] Traditional machine learning methods (XGBoost, CatBoost, LightGBM) dominatetabular data benchmarks but lack explicit uncertainty quantification. Recent deeplearning approaches for tabular data (TabNet, FT-Transformer) introduce attentionmechanisms but suffer from critical limitations:a. Feature-level attention only- lacking sample-level filtering to identify uncertainpredictions.b. Limited statistical context- do not incorporate correlation and covariancestructures.c. No uncertainty quantification- provide no probabilistic measure of samplereliability.d. High computational requirements- unsuitable for edge device deployment.e. Poor noise robustness- fail on datasets with outliers and imbalanced classes.Uma-Attention addresses all these limitations through information-theoreticfoundations, dual-level filtering, computational efficiency, and explicit noisehandling.[4] Core Technical Problem that exists is how can a learning model automatically filterboth unreliable features and uncertain samples in data while remaining computationallyefficient and deployable on edge devices. Data represents the most prevalent dataformat in enterprise and scientific applications. Existing deep learning solutionsstruggle with multiple challenges:1. Noise and Outlier Sensitivity: Datasetsfrequently contain outliers, missing values, andnoisy samples. Existing attention mechanismsdo not explicitly identify or downweightuncertain samples.2. 2 Uma-Attention: Feature-Sample GranularityMismatch: Most existing mechanisms (SE-Net,channel attention) operate at feature level only.They do not distinguish between highconfidence and low confidence predictions atthe sample level.3. Statistical Context Loss: Traditional gatingnetworks fail to leverage correlation andcovariance structures inherent in data.4. Edge Deployment Constraints: State-of-the-artmethods (transformers, large ensemble models)require extensive computation unsuitable forresource-constrained IoT and edge devices.5. Imbalanced Data Handling: In fraud detection,medical diagnostics, and anomaly detection,imbalanced datasets cause standard deeplearning approaches to fail. Existing attentionmechanisms provide no mechanism toexplicitly handle sample uncertainty[5] A conventional solution discloses channel attention via global pooling and MLP-basedgating . Applied originally to images, now adapted for data. Limitation: Operates onlyat feature / channel level; uses only pooling without statistical context.[6] Another conventional solution discloses an expert routing and conditional computationenabling parallel processing. Scalable MoE approaches for conditional computation.Limitation: Lacks explicit statistical context in routing decisions; no sample-levelfiltering.[7] Another conventional solution discloses a sequential sparse feature masking withdecision-tree style attention. Provides feature selection interpretability. Limitation:Lacks sample-level filtering; limited noise robustness.[8] Another conventional solution discloses an adaptation of transformer architectures withembeddings for categorical features. Limitation: High computational cost; no samplelevel filtering; unsuitable for edge deployment.[9] Another conventional solution discloses a tree-inspired neural approaches combiningdecision trees and neural networks. Limitation: Not designed for noise-robust samplefiltering; limited theoretical justification.

[10] Yet another conventional solution discloses a multi-stage attention in residualconnections. Limitation: Designed for images; different data characteristics comparedto data.

[11] Limitations of Academic Methods • SE-Net (Squeeze-and-ExcitationNetworks):

[12] Channel attention via global pooling discards instance-specific information.Cannot distinguish high-confidence from low-confidence samples. No incorporation offeature correlation structure. Limited to image domain with spatial structureassumptions.

[13] TabNet(Attentive Interpretable Tabular Learning): Provides interpretablefeature mask ing but lacks explicit sample certainty filtering. Sequential attention overfeatures alone proves insufficient for data with high noise or uncertainty. No theoreticalfoundation (e.g., information theoretic principles) justifying attention mechanismdesign. Requires large parameter counts (500K+) limiting edge deployment.

[14] Transformers (Attention is All You Need): Multi-head self-attention operateson sequential data with quadratic complexity (O(n2)). No mechanism to identify ordown weight uncertain samples. No explicit feature correlation modelling for data.Unsuitable for edge deployment due to high memory requirements. Positionembeddings and sequential assumptions misaligned with data structure (which has noinherent ordering).

[15] FT-Transformer (Feature Tokenizer Transformer): Addresses data butmaintains quadratic attention complexity. Embedding layer overhead increasesparameter count. No sample level confidence / uncertainty mechanism. Lacks theoreticalgrounding in information theory for tabular-specific filtering.

[16] Mixture-of-Experts (MoE) Methods: Routing decisions based on learnedassignments with out statistical context. Does not leverage correlation structures infeature space. No sample-level filtering for uncertainty handling. Generic gating designnot optimized for data characteristics.

[17] NODE (Neural Oblivious Decision Ensembles): Tree-inspired architecture butlimited flexibility in attention mechanism design. Not specifically optimized forfeature / sample filtering trade-offs. Requires retraining for different data characteristicsor domains.

[18] A conventional solution focuses on physical engineering systems rather thanmachine learning. Generic edge computing optimization without neural architecturespecialization for data. No information-theoretic mechanisms for data filtering.

[19] Another conventional solution discloses a domain-specific to communicationsystems. Lacks explicit sample-level attention mechanism. No documented featurecorrelation modelling. Limited generalizability to other domains.

[20] Another conventional solution discloses a conceptually similar but high-levelwith out specific technical implementation. Uses biological metaphors rather thanmathematical foundations. No concrete mathematical formulation for attentionmechanisms.

[21] Absence of Sample-Level Filtering: Existing methods focus exclusively onfeature-level attention. No prior work applies Shannon entropy-based weighting at thesample level for uncertainty quantification and noise suppression in tabularclassification.

[22] Missing Statistical Context: Feature-level gating mechanisms (SE-Net, TabNet,attention blocks) do not incorporate feature covariance or correlation structurestatistics. Uma-Attention augments gating with correlation energy, providing contextaware feature importance.

[23] Lack of Information-Theoretic Foundation: Attention mechanisms typicallyrely on learned weights without theoretical justification. Uma-Attention grounds bothsample and feature filtering in Shannon entropy principles from information theory.

[24] Edge Deployment Trade-offs: Prior solutions either achieve high accuracy withlarge models (TabNet: 500K parameters) or sacrifice performance for efficiency. UmaAttention achieves 85.5% accuracy with only 3,926 parameters, making it genuinelydeployable on resource-constrained devices.

[25] Preprocessing Burden: Most methods require careful hyperparameter tuning ofpreprocessing (outlier thresholds, imputation strategies, feature scaling). UmaAttention requires only minimal preprocessing (missing value removal, categoricalencoding, standardization)

[26] Tabular Data Specificity: Transformers and CNNs have implicit data structureassumptions (sequential dependencies, spatial structure). Uma-Attention isarchitecturally designed for tabular data: independent feature dimensions with mixeddata types.

[27] There is a need to overcome the above-mentioned drawbacks.OBJECTS OF THE INVENTION

[28] Some of the objects of the present disclosure, which at least one embodimentherein satisfy, are listed herein below.

[29] The primary object of the present subject matter is to provide a method forclassifying structured data.

[30] Another object of the present subject matter is to provide a dual-level filtering:Implement complementary feature-level (via Correlation-Excitation) and sample-level(via Adaptive Entropy Excitation) filtering to suppress both feature noise and uncertainpredictions.

[31] Another object of the present subject matter is to provide ground attentionmechanisms in Shannon entropy and information theory principles rather thanheuristics, providing theoretical justification for design choices.

[32] Another object of the present subject matter is to augment neural network gatingwith correlation energy statis tics to leverage inter-feature dependencies specific todata.

[33] Another object of the present subject matter is to enable explicit handling ofoutliers, imbalanced data, and noisy samples through entropy-based sample uncertaintyquantification.

[34] Another object of the present subject matter is to design architecture suitable foredge computing deployment on resource-constrained IoT devices while maintainingcompetitive performance.

[35] Another object of the present subject matter is to provide interpretable attentionweights (entropy values indicate sample confidence; correlation energy indicatesfeature activation patterns) enabling human understanding of model decisions.

[36] Another object of the present subject matter is to ensure full differentiability ofall components for gradient-based optimization and integration with standard deeplearning frameworks.

[37] Another object of the present subject matter is to design components asindependent building blocks that can be used in different combinations for diverseapplications.

[38] This and other objects and advantages will become more apparent whenreference is made to the following description and accompanying drawings.SUMMARY OF THE INVENTION

[39] This summary is provided to introduce concepts related to a method forclassifying data. The concepts are further described below in the detailed description.This summary is not intended to identify key features or essential features of theclaimed subject matter, nor is it intended to be used to limit the scope of the claimedsubject matter.

[40] The present subject matter discloses a method for classifying data. The methodincludes receiving, by a receiving engine, a plurality of feature vectors associated withdata as an input. The method includes performing, by a partitioning engine, an adaptivestream partitioning of the plurality of features into a plurality of streams. The methodincludes processing, by a parallel processing engine, the plurality of streams parallellyto one another to generate a plurality of processed streams. The method includesperforming, by a fusion engine, a sparse adaptive fusion on the plurality of processedstreams to concatenate the plurality of processed streams into a fused vector. Themethod includes classifying, by a classification engine, the data based into one or moreclasses based on a probability assigned to each stream in the fused vector

[41] The present subject matter discloses a system for classifying data. The systemincludes receiving, by a receiving engine, a plurality of feature vectors associated withdata as an input. The system includes performing, by a partitioning engine, an adaptivestream partitioning of the plurality of features into a plurality of streams. The systemincludes processing, by a parallel processing engine, the plurality of streams parallellyto one another to generate a plurality of processed streams. The system includesperforming, by a fusion engine, a sparse adaptive fusion on the plurality of processedstreams to concatenate the plurality of processed streams into a fused vector. Thesystem includes classifying, by a classification engine, the data based into one or moreclasses based on a probability assigned to each stream in the fused vector

[42] Various objects, features, aspects, and advantages of the inventive subjectmatter will become more apparent from the following detailed description of preferredembodiments, along with the accompanying drawing figures in which like numeralsrepresent like components.BRIEF DESCRIPTION OF THE ACCOMPANYING DRAWINGS

[43] The illustrated embodiments of the subject matter will be understood byreference to the drawings, wherein like parts are designated by like numeralsthroughout. The following description is intended only by way of example and simplyillustrates certain selected embodiments of devices, systems, and methods that areconsistent with the subject matter as claimed herein, wherein:

[44] Fig. 1 illustrates a diagram depicting an environment comprising a system forclassifying data, in accordance with an embodiment of the present subject matter;

[45] Fig. 2 illustrates a schematic block diagram depicting a system for classifying data, inaccordance with an embodiment of the present subject matter;

[46] Fig. 3 illustrates an operational flow diagram depicting a process for classifyingdata, in accordance with an embodiment of the present subject matter;

[47] Fig. 4 illustrates an operational flow diagram depicting a process for performingan adaptive stream partitioning of data, in accordance with an embodiment of thepresent subject matter;

[48] Fig. 5 illustrates an operational flow diagram depicting a parallel processing ofa number of streams, in accordance with an embodiment of the present subject matter;

[49] Fig. 6 illustrates an operational flow diagram depicting a sparse adaptive fusion,in accordance with an embodiment of the present subject matter;

[50] Fig. 7 illustrates an operational flow diagram depicting a classification head, inaccordance with an embodiment of the present subject matter;

[51] Fig. 8a and 8b illustrate line diagrams depicting use cases of the system onphysiological signals with processing and without processing, in accordance with anembodiment of the present subject matter; and

[52] Fig. 9 illustrates a schematic block diagram depicting a method for classifyingdata, in accordance with an embodiment of the present subject matter.

[53] The figures depict embodiments of the present subject matter for the purposesof illustration only. A person skilled in the art will easily recognize from the followingdescription that alternative embodiments of the structures and methods illustratedherein may be employed without departing from the principles of the disclosuredescribed herein.DETAILED DESCRIPTION

[54] The following is a detailed description of embodiments of the disclosuredepicted in the accompanying drawings. The embodiments are in such detail as toclearly communicate the disclosure. However, the amount of detail offered is notintended to limit the anticipated variations of embodiments; on the contrary, theintention is to cover all modifications, equivalents, and alternatives falling within thespirit and scope of the present disclosure as defined by the appended claims.

[55] While the embodiments of the disclosure are subject to various modificationsand alternative forms, specific embodiments thereof have been shown by way ofexample in the figures and will be described below. It should be understood, however,that it is not intended to limit the disclosure to the particular forms disclosed, but on thecontrary, the disclosure is to cover all modifications, equivalents, and alternativesfalling within the scope of the disclosure.

[56] The terms "comprises", "comprising", or any other variations thereof used inthe disclosure, are intended to cover a non-exclusive inclusion, such that a device,system, or assembly that comprises a list of components does not include only thosecomponents but may include other components not expressly listed or inherent to suchsystem, or assembly, or device. In other words, one or more elements in a system ordevice proceeded by "comprises… a" does not, without more constraints, preclude theexistence of other elements or additional elements in the system or device.

[57] Fig. 1 illustrates a diagram depicting an environment 100 comprising a system102 for classifying data, in accordance with an embodiment of the present subjectmatter. The data may be one of image data, video data, audio data, textual data, tabulardata, sequential data, and a combination of at least two of the image data, the videodata, the audio data, the textual data, the tabular data, and the sequential data beingclassified in a neural network architecture. The system 102 may be optimized forstructured data where instances are represented as fixed-dimensional feature vectors.Data exhibits strong inter-feature correlations. By providing a gating network withcorrelation statistics, it learns to model dependencies between features, enabling moreinformed gating decisions. The system 102 may be configured to incorporate statisticalcontext beyond raw features, enable learning of feature interaction patterns, leverageinherent data structure, and maintain computational efficiency through low-rankbottleneck design.

[58] In accordance with an embodiment of the present subject matter, the system 102may be configured to receive a number of feature vectors associated with the data as aninput. Upon receiving the number of feature vectors, the system 102 may be configuredto perform an adaptive stream partitioning of the number of feature vectors into anumber of streams. The adaptive stream partitioning may route the number of featurevectors to parallel processing streams based on learned softmax assignments, enablingspecialization of different streams to different feature subsets. The number of featurevectors may be related to features of the data. Specifically, the features may be shape(batch_size, input_features).

[59] Moving forward, upon a generation of the number of streams, the system 102may be configured to process the number of streams parallelly to one another. Thenumber of streams may be processed parallelly to generate a number of processedstreams. The parallel processing of the number of streams may include a DynamicFeature Energy Normalization (DFEN), a dense projection layer processing, aCorrelation-Excitation (CE) block for feature gating, and an Adaptive EntropyExcitation (AEE) layer for sample filtering. The system 102 may be a modulararchitecture enabling an Uma-Attention to achieve high accuracy with minimalparameters through information-theoretic filtering (AEE), correlation-aware gating(CE), and adaptive stream spe cialization (ASP+SAF).

[60] In response to parallelly processing the number of streams, the system 102 maybe configured to perform a sparse adaptive fusion on the number of processed streamsto concatenate the number of processed streams into a fused vector. Upon generationof the fused vector, the system 102 may be configured to classify the data based intoone or more classes based on a probability assigned to each stream in the fused vector.

[61] Fig. 2 illustrates a schematic block diagram 200 depicting the system 102 forclassifying data, in accordance with an embodiment of the present subject matter. Thedata may be one of image data, video data, audio data, textual data, tabular data,sequential data, and a combination of at least two of the image data, the video data, theaudio data, the textual data, the tabular data, and the sequential data.

[62] Continuing with the above embodiment, the system 102 may include aprocessor 202, a memory 204, data 206, a receiving engine 208, a pre-processing engine210, a partitioning engine 212, a parallel processing engine 214, a fusion engine 216,and a classification engine 218. The processor 202, a memory 204, data 206, thereceiving engine 208, the pre-processing engine 210, the partitioning engine 212, theparallel processing engine 214, the fusion engine 216, and the classification engine 218may be communicatively coupled to one another.

[63] In an aspect, the processor 202 may be implemented as one or moremicroprocessors, microcomputers, microcontrollers, digital signal processors, centralprocessing units, logic circuitries, and / or any devices that manipulate data based onoperational instructions. Among other capabilities, the processor 202 may beconfigured to fetch and execute computer-readable instructions stored in the memory204 of the systems 102.

[64] In an aspect, the processor 202 may be implemented as a combination ofhardware and programming device(s) (for example, programmable instructions) toimplement one or more functionalities of the processing device. In examples describedherein, such combinations of hardware and programming may be implemented inseveral different ways. In one example, the programming for the processor 202 may beprocessor executable instructions stored on a non-transitory machine-readable storagemedium and the hardware for the processor 202 may include a processing resource (forexample, one or more processors), to execute such instructions. In other examples, theprocessor 202 may be implemented by an electronic circuitry.

[65] Further, in an aspect, the memory 204 may store one or more computer-readableinstructions or routines, which may be fetched and executed to create or share data unitsover a network service. The memory 204 may include any non-transitory storage deviceincluding, for example, volatile memory such as RAM, or non-volatile memory suchas EPROM, flash memory, and the like.

[66] In an aspect, the memory 204 may include any non-transitory computerreadable medium known in the art including, for example, volatile memory, such asstatic random-access memory (SRAM) and / or dynamic random-access memory(DRAM), and / or non-volatile memory, such as read-only memory (ROM), erasableprogrammable ROM (EPROM), flash memory, hard disks, optical disks, and / ormagnetic tapes. The memory 204 may further include the data 206.

[67] The data 212 serves, amongst other things, as a repository for storing dataprocessed, received, and generated by the system 102.

[68] Continuing with the above embodiment, the receiving engine 208 may beconfigured to receive an input. The input may be the data. The pre-processing engine210 may be configured to pre-process the data to generate a number of feature vectorsassociated with the data. The pre-processing engine 210 may be configured to transmitthe number of feature vectors to the partitioning engine 212.

[69] Upon a successful receipt of the input by the partitioning engine 212, thepartitioning engine 212 may be configured to perform an adaptive stream partitioningof the number of feature vectors. The number of feature vectors may be partitioned intoa number of streams. For performing the adaptive stream partitioning, the partitioningengine 212 may be configured to generate a probability matrix associated with thenumber of feature vectors. The probability matrix may be generated via a softmaxrouting. The probability matrix may include a number of probabilities. Upon thegeneration of the probability matrix, the partitioning engine 212 may be configured toprocess the number of feature vectors with each probability in the probability togenerate the number of streams.

[70] Upon a partition of the number of feature vectors into a number of streams bythe partitioning engine 212, the parallel processing engine 214 maybe configured toprocess the number of streams parallelly to one another. The number of streams maybe processed in order to generate a number of processed streams. For processing thenumber of streams parallelly, the parallel processing engine 214 may be configured toperform a normalization of each streams amongst the number of streams to stabilize afeature magnitude associated with each stream. Upon performing the normalization, theparallel processing engine 214 may be configured to process the number of streamshaving a stabilized feature magnitude via a ReLU activation. Upon the processing, theparallel processing engine 214 may be configured to augment each stream withcorrelation energy statistics and apply a learned gating on each stream.

[71] Data exhibits strong inter-feature correlations. By providing the gating networkwith correlation statistics, it learns to model dependencies between features, enablingmore informed gating decisions. Further, the parallel processing engine 214 may beconfigured to compute a normalized Shannon entropy associated with each stream togenerate the number of processed streams. The normalized Shannon entropy may be adifferentiable sample-level attention mechanism for neural networks. Standardattention mechanisms operate at feature level only. The parallel processing engine 214may perform a sample-level filtering by computing Shannon entropy from normalizedinput feature values:

[72] Low entropy indicates concentrated feature activation (few features dominate),suggesting high prediction certainty. High entropy indicates uniform activation (allfeatures equally important), suggesting low certainty or noise. By computing w =max(0, min(1, 1-Hnorm)) where Hnorm = H / log(n), uncertain samples receive lowerweights, effectively filtering noisy predictions.

[73] Further, the parallel processing engine 214 may be configured to calculate anattention weight associated with each stream based on the normalized Shannon entropy.Upon calculating the attention weight, the parallel processing engine 214 may beconfigured to scale the number of streams with respect to the attention weight togenerate the number of processed streams. To that understanding, the parallelprocessing engine 214 may be configured to compute the normalized Shannon entropyby computing a Shannon entropy associated with each stream, and normalizing theShannon entropy with a maximum Shannon entropy to generate the normalizedShannon entropy. Normalizing the Shannon entropy may be performed to reduce noisefrom the number of streams.

[74] Normalize feature absolute values to probability distribution:

[75] Compute Shannon entropy:

[76] Normalize by maximum entropy:

[77] Compute attention weight as complement of normalized entropy:

[78] Apply scaling to input:

[79] The fusion engine 216 may be configured to perform a sparse adaptive fusionon the number of processed streams. The sparse adaptive fusion may be performed onthe number of processed streams for concatenating the number of processed streamsinto a fused vector. For performing the sparse adaptive fusion, the fusion engine 216maybe configured to concatenate the number of streams with one another to generate avector. Upon generating the vector, the fusion engine 216 may be configured to removea noise from the fused vector by performing a dimensionality reduction on the vector.A trust score is assigned to each stream in the fused vector upon the noise removal.Further, the fusion engine 216 may be configured to fuse the vector with each trustscore associated with number of streams to generate the fused vector.

[80] Upon a generation of the fused vector by the fusion engine 216, theclassification engine 218 may be configured to classify the data based into one or moreclasses based on a probability assigned to each stream in the fused vector. Forperforming the classification, the classification engine 218 may be configured toperform a batch normalization of the fused vector to normalize a number of activationsacross a batch dimension. Upon performing the batch normalization, the classificationengine 218 may be configured to randomly dropping a predetermined percentage of thenumber of activation. Furthermore, the classification engine 218 may be configured togenerate a probability associated with the data belonging to the one or more classes.

[81] Fig. 3 illustrates an operational flow diagram depicting a process 300 forclassifying data, in accordance with an embodiment of the present subject matter. Theprocess 300 may be performed by the system as referred in fig. 1 and fig. 2. The process300 may be implemented in a neural network architecture and the data may be one ofimage data, video data, audio data, textual data, tabular data, sequential data, and acombination of at least two of the image data, the video data, the audio data, the textualdata, the tabular data, and the sequential data. The process 300 may combine a numberof features to classify the data.

[82] At step 302, the process 300 may include receiving an input. The input maybereceiving by the partitioning engine 212 and the input may be a number of featurevectors associated with the data.

[83] At step 304, the process 300 may performing an adaptive stream partitioning ofthe number of feature vectors. The number of feature vectors may be partitioned into anumber of streams. The adaptive stream partitioning may be performed by thepartitioning engine 212. For performing the adaptive stream partitioning, thepartitioning engine 212 may be configured to generate a probability matrix associatedwith the number of feature vectors. The probability matrix may be generated via asoftmax routing. The probability matrix may include a number of probabilities. Uponthe generation of the probability matrix, the partitioning engine 212 may be configuredto process the number of feature vectors with each probability in the probability togenerate the number of streams.

[84] At step 306, the process 300 may include performing a normalization of eachstreams amongst the number of streams to stabilize a feature magnitude associated witheach stream. The normalization may be performed by the parallel processing engine214 as a first step of a four step processing of the number of streams.

[85] At step 308, the process 300 may include processing the number of streamshaving a stabilized feature magnitude via a ReLU activation. The processing may beperformed by the parallel processing engine 214 as a second step of the four stepprocessing of the number of streams.

[86] At step 310, the process 300 may include augmenting each stream withcorrelation energy statistics and apply a learned gating on each stream. The augmentingmay be performed by the parallel processing engine 214 as a third step of the four stepprocessing of the number of streams.

[87] Mean-center input features:

[88] Compute correlation energy:

[89] Augment original features with correlation energy:

[90] Dense bottleneck network with reduction ratio r = 2:

[91] Sigmoid gating network:

[92] Element-wise multiplication:

[93] At step 312, the process 300 may include computing a normalized Shannonentropy associated with each stream to generate the number of processed streams. Thecomputation may be performed by the parallel processing engine 214 as a fourth stepof the four step processing of the number of streams. Further, the parallel processingengine 214 may be configured to calculate an attention weight associated with eachstream based on the normalized Shannon entropy. Upon calculating the attentionweight, the parallel processing engine 214 may be configured to scale the number ofstreams with respect to the attention weight to generate the number of processedstreams. To that understanding, the parallel processing engine 214 may be configuredto compute the normalized Shannon entropy by computing a Shannon entropyassociated with each stream, and normalizing the Shannon entropy with a maximumShannon entropy to generate the normalized Shannon entropy. Normalizing theShannon entropy may be performed to reduce noise from the number of streams. Thenumber of processed streams may be used as an input in the fusion engine 216.

[94] At step 314, the process 300 may include performing a sparse adaptive fusionon the number of processed streams. The sparse adaptive fusion may be performed onthe number of processed streams for concatenating the number of processed streamsinto a fused vector by the fusing engine. For performing the sparse adaptive fusion, thefusion engine 216 maybe configured to concatenate the number of streams with oneanother to generate a vector. Upon generating the vector, the fusion engine 216 may beconfigured to remove a noise from the fused vector by performing a dimensionalityreduction on the vector. A trust score is assigned to each stream in the fused vectorupon the noise removal. Further, the fusion engine 216 may be configured to fuse thevector with each trust score associated with number of streams to generate the fusedvector.

[95] At step 316, the process 300 may include classifying the data into one or moreclasses based on a probability assigned to each stream in the fused vector. Forperforming the classification, the classification engine 218 may be configured toperform a batch normalization of the fused vector to normalize a number of activationsacross a batch dimension. Upon performing the batch normalization, the classificationengine 218 may be configured to randomly dropping a predetermined percentage of thenumber of activation. Furthermore, the classification engine 218 may be configured togenerate a probability associated with the data belonging to the one or more classes.

[96] Fig. 4 illustrates an operational flow diagram 400 depicting a process forperforming an adaptive stream partitioning of data, in accordance with an embodimentof the present subject matter. The ASP may be performed to split a number of featurevectors related to the data into a number of streams. The ASP may be performed by thepartitioning engine 212 as referred in fig. 2.

[97] The partitioning engine 212 may learn to route the number of feature vectors toS parallel streams via a learned softmax assignment. The number of feature vectors maybe of a shape (BF) where B is batch size, F is a number of features related to the data.The ASP may include an assignment layer that may be a dense layer with softmaxactivation projecting features to stream assignment probabilities.A. Operation: Dense(F--> S) + SoftmaxB. Output shape: (BS) where each sample gets S-dimensionalprobability distribution

[98] Further, the ASP may include performing a weighted distribution. The weighteddistribution may include an Einstein summation that computes weighted featureassignment. The weighted feature assignment may be computed based on einsum ('bi,bj-> bij', input_features, assignment). Further, S weighted copies of input features maybe created and an output shape may be SX(BF)- one tensor per stream. This allowseach stream to specialize in different feature subsets or patterns, enabling mixture-ofexperts style parallel processing.

[99] Fig. 5 illustrates an operational flow diagram 500 depicting a parallel processingof a number of streams, in accordance with an embodiment of the present subjectmatter. The number of streams may be generated from a number of feature vectors ofdata. The data may be one of image data, video data, audio data, textual data, tabulardata, sequential data, and a combination of at least two of the image data, the videodata, the audio data, the textual data, the tabular data, and the sequential data to beclassified into one or more classes. The parallel processing may be performed by theparallel processing engine 214 as referred in Fig. 2.

[100] In parallel processing, the number of streams "S" processes it's input throughan identical pipeline. The input for the pipeline may be weighted feature tensor (BF)from the partitioning engine 212. The weighted feature tensor (BF) may be one of thenumber of streams. The parallel processing of the number of streams may include aDynamic Feature Energy Normalization (DFEN), a dense projection layer processing,a Correlation-Excitation (CE) block for feature gating, and an Adaptive EntropyExcitation (AEE) layer for sample filtering.

[101] The DFEN may include applying L2 normalization on the input to stabilizefeatures magnitude across samples. Shape (B,F) may be an output.

[102] Further, the dense projection layer processing (Dense(F ->P, activation=ReLU))may take place on each stream upon normalization. It may include projecting tointermediate dimension P (typically 12 for Adult UCI) to give an output shape (B, P).Continuing with the above embodiment, the CE block may include performing a meancentering (xc = x-mean(x, axis = 0)). Upon the mean centering, a correlation energymay be calculated based on:

[103] The CE block augments traditional channel attention with correlation energystatistics to capture feature covariance structures. Further, a concatenation [x,Ecorr] maybe performed. Upon the concatenation, a bottleneck gating may be performed based onthe below formula:

[104] Upon bottleneck gating, a feature gating may be performed which may includemultiplying original features with learned gate element wise to provide an output (B,P).Furthermore, the AEE block may include performing an entropy calculation where anormalized Shannon entropy may be computed from feature activations. Upon thecomputation of the normalized Shannon entropy, a weight computation may beperformed, w = 1.0-Hnorm (inverse entropy as attention weight). Further, a sample levelmultiplication may be performed on scale features based on entropy weight. To give anoutput (B,P). This final output may be a number of processed streams that may be fusedahead. This pipeline is replicated across all S streams (typically S = 4), producing Sindependent processed tensors. The AEE layer computes normalized Shannon entropyfrom input features to identify uncertain samples, down weighting those with uniformfeature activation patterns.

[105] Fig. 6 illustrates an operational flow diagram 600 depicting a sparse adaptivefusion, in accordance with an embodiment of the present subject matter. The sparseadaptive fusion may be performed by the fusion engine 216 as referred in fig. 2 to fusea number of processed streams with one another. The number of processed streams maybe of shape (B, P). The number of processed streams may be concatenated along featuredimension to provide an output shape (B, S X P). In an exemplary embodiment, S = 4,P =12: (B48). Further, a low rank bottleneck may be performed that may includedimensionality reduction for efficient gating computation (Dense (S X P-> (S X P) / 2,activation=ReLU) to provide an output shape (B, LowRank).

[106] To that understanding, a gating network may further perform a dense layerprocessing on the number of processes streams (Dense (LowRank->S X P,activation=Sigmoid)), project the number processed streams back to concatenateddimension and outputs element wise gate values ε [0, 1]. An output may be shape (B, SX P). Moving forward, an element wise gating may further be performed whereconcatenated features may be multiplied with learned weights. An output may be fusedvector of shape (B, S X P).

[107] The low-rank gating mechanism learns which stream outputs (or streamcombinations) to emphasize for each sample, enabling adaptive fusion based on inputcharacteristics.

[108] Fig. 7 illustrates an operational flow diagram 700 depicting a classificationhead, in accordance with an embodiment of the present subject matter. Theclassification head may be configured to classify data associated with a fused vectorinto one or more classes based on generating a probability for the data belonging toeach of the one or more classes. The classification may be performed by theclassification engine 218 as referred in Fig. 2. The classification engine 218 may receivethe fused vector of shape (B, S X P) as an input. Further, the classification engine 218may perform a batch normalization on the input to normalize activations across batchdimensions. The batch normalization may stabilise training and improve convergence.An output shape (B, S X P) may be provided.

[109] Further, the classification engine 218 may be configured to randomly drop apercentage of activations during training (typically p = 50%) to prevent overfitting byforcing network redundancy and provide an output shape (B, S X P). To thatunderstanding, a softmax operation may be applied (Dense (SXP->C,activation=Softmax)) and the fused vector is projected to C classes. The softmaxoperation ensures output probabilities sum to 1.0. The output shape is (B, C). Eachsample receives a C-dimensional probability distribution and a predicted class with amaximum confidence score is mentioned below.

[110] Fig. 8a and 8b illustrate line diagrams 800a, and 800b depicting use cases of thesystem 102 on physiological signals with processing and without processing, inaccordance with an embodiment of the present subject matter. The system 102 may beemployed in an adult UCI Dataset Full Pipeline.

[111] Problem: Binary classification to predict whether an individual's annual incomeexceeds $50,000 based on demographic and employment features. Initially, a dataacquisition and an initial loading takes place. Raw CSV data is downloaded directlyfrom UCI Machine Learning Repository. Dataset contains 48,842 instances with 14input features and 1 binary target. Features include age, workclass, education, maritalstatus, occupation, relationship, race, sex, capital-gain, capital-loss, hours-per-week,native-country, etc. Target variable may be income ∈ {<=50K, >50K}. Missing valuesare marked with ' ?' character.

[112] The system 102 may be configured to perform a pre-processing. The preprocessing may include a missing value handling. Upon an identification of missing orundefined entries within a dataset, the preprocessing engine may be configured toperform a number of corrective operations to address such missing information. In oneexample, the preprocessing engine may be configured to remove an entire data recordcontaining a missing entry so as to maintain the integrity of the remaining dataset. Inanother example, the preprocessing engine may be further configured to replace themissing entry with a representative statistical value, such as a mean, median, mode, oranother imputed estimate derived from the available data. Upon performing one or moreof these corrective operations, a refined dataset may be generated for subsequentprocessing. The system 102 may further be configured to perform a categorical featureencoding. Upon determining that the dataset includes a number of categorical attributesrepresented in textual form, an encoding module may be configured to transform eachattribute into a corresponding numerical representation. The encoding module may befurther configured to assign a distinct numerical identifier to each unique categoricalvalue so as to generate a consistent numerical format across all samples. A similartransformation may be applied to a binary output label to convert it into numerical form.Upon performing the transformation, the dataset may comprise a set of numericalfeatures suitable for mathematical modeling.Moving forward, the system102 may be configured to perform a featurestandardization. Upon a partition of the dataset into training, validation, and testsubsets, a normalization engine may be configured to perform a stabilization of featuremagnitudes by generating a standardized representation of each numerical attribute.The normalization engine may be further configured to compute a central tendency andvariability measure using only the training subset so as to avoid statistical leakage.Upon computing these measures, the normalization engine may apply a scalingoperation to all subsets such that each feature exhibits a normalized magnitudeappropriate for downstream processing.

[113] Further, the data may be split into a training set, a validation set and a test set.Training set may include 20,838 samples (64% of total), for model parameter learning.The validation set may include 5,212 samples (16% of total) for hyperparameter tuningand early stopping. The test set may include 6,513 samples (20% of total) for finalunbiased performance evaluation. Stratified splitting ensures class balance preservationacross all subsets. Class distribution may be ≈75.9% (<=50K) vs. ≈24.1% (>50K). Thesystem 102 may tune a number of hyperparameters. The number of hyperparametersmay include:a. Number of parallel streams (nstreams): 4b. Projection size per stream: 12c. Dropout rate: 0.5 (50% dropout for regularization)d. Learning rate: 1.0 x10-3 (Adam optimizer)e. Total trainable parameters: 3,926 (15.34 KB model size).

[114] Further, the system 102 may receive 14-dimensional feature vectors associatedwith the data to classify the data. The system 102 may be undergo a training and thetraining configuration may include:a. Optimizer: Adam (default parameters: β1 = 0.9, β2 = 0.999)b. Loss function: Sparse categorical cross-entropyc. Batch size: 32 samples per mini-batchd. Maximum epochs: 200 (early stopping prevents overfitting)e. Callbacks:1. EarlyStopping: Monitors validation loss,patience = 15 epochs, restores best weights.2. ReduceLROnPlateau: Reduces learning rate byfactor 0.5 when validation loss plateaus (patience = 5)

[115] Training progression may include:a. Epoch 1: Training accuracy = 77.90%, Validation accuracy = 82.05%b. Epoch 5: Training accuracy = 84.36%, Validation accuracy = 84.45%c. Epoch 18 (Best): Training accuracy = 85.00%, Validation accuracy = 84.61%(best validation loss = 0.3282)d. Epoch 33: Early stopping triggered after 15 epochs without improvemente. Total training time: 1.01 minutes on CPUf. Learning rate reduced 4 times during training: 10-3 -> 5 x 10-4 -> 2.5 x 10-4 ->1.25x10-4 -> 6.25 x 10-5

[116] After training completion, the model (with weights restored to best validationepoch) is evaluated on the held-out test set:

[117] Further, a confusion matrix analysis may include:a. True Negatives (<=50K correctly predicted): 4,669b. True Positives (>50K correctly predicted): 900c. False Positives (<=50K incorrectly predicted as >50K): 276d. False Negatives (>50K incorrectly predicted as <=50K): 668

[118] The system 102 may generate following output:a. Trained Model: Saved weights file containing learned parameters (15.34 KB)b. Predictions: Class probabilities for all 6,513 test samplesc. Performance Metrics: Accuracy (85.51%), AUC-ROC (0.9086), precision,recall, F1-scoresd. Confusion Matrix Visualization: Heatmap showing classification performancebreakdowne. ROC Curve: Plot of True Positive Rate vs. False Positive Rate across decisionthresholdsf. Training History Plots:1. Loss curve: Training loss vs. validation lossover 33 epochs2. Accuracy curve: Training accuracy vs.validation accuracy over 33 epochs

[119] Key Results may include:a. Compact Architecture: Only 3,926 trainable parameters (125x fewer thanTabNet's 500,000)b. High Accuracy: 85.51% test accuracy competitive with state-of-the-art tabularmethodsc. Strong Discrimination: AUC-ROC = 0.9086 indicates excellent class separationd. Fast Training: Convergence in 33 epochs (1 minute on CPU)e. Edge Deployable: 15.34 KB model suitable for resource-constrainedenvironmentsf. Minimal Preprocessing: No outlier removal, no imputation, no featureengineering requiredg. Robust Handling: Successful performance despite class imbalance (75.9% vs.24.1%)

[120] Above example demonstrates an effectiveness of the system 102 as a generalpurpose data classifier, combining information-theoretic principles with practicaldeployment efficiency.

[121] In another embodiment, the system 102 may be a part of a wearable healthmonitoring system for continuous stress detection in hospital patients, combining edgeinference with cloud analytics. The wearable health monitoring system may include anedge device such as a microcontroller-based wearable sensor module, a number ofsensors to sense a heart rate (PPG) sensor, skin temperature, electrodermal activity(EDA), and accelerometer, a backend server, and the system 102. The edge device maybe configured to quantify the input from 32-bit float to 8-bit integer weights with anaccuracy degradation of less than 1 %.

[122] The micro-controller may read a number of features. The number of featuresinclude Heart rate, HRV (RMSSD), respiratory rate, skin temperature, EDA,accelerometer X / Y / Z. Raw value of the number of features are normalized using precomputed statistics. The system 102 receives 8-dimensional feature vectors associatedwith normalized raw values. An output is generated amongst 3 class probabilities(Baseline, Stress, Amusement). Further, the system 102 estimates a latency of 12 msand energy of 0.267 mJ. If stress is detected, the system 102 may trigger a local alert.The output is stored in a local buffer. The system 102 monitors the feature vectorsindependent of a network continuously.

[123] Further, when a wireless network such as Wi-Fi is available, the back end servermay perform a synchronization. Buffered predictions are batch uploaded to AWS IoTCore (MQTT), cloud lambda function aggregates data for trend analysis. Medical staffis alerted based on configurable criteria.

[124] Verified Performance Metrics (from Adult UCI Implementation):

[125] Estimated Deployment Characteristics (Pending Validation):

[126] Future Validation Required:a. On-device quantization and inference timing measurements on actual ESP32hardwareb. Physiological feature extraction and preprocessing pipeline implementationc. Battery life validation under continuous monitoring conditionsd. Accuracy retention after quantization on WESAD and patient stress detectiondatasetse. Cloud integration and MQTT message throughput testing

[127] A number of advantages (design-level):a. Privacy: Architecture supports local-first data processing; cloud transmissionoptionalb. Offline Capability: Inference possible without network connectivityc. Parameter Efficiency: 3,926 parameters enables edge deployment compatibilityd. Scalability: Parallel-stream design allows multiple device deployment

[128] Fig. 9 illustrates a schematic block diagram depicting a method 900 forclassifying data, in accordance with an embodiment of the present subject matter. Themethod 900 may be performed by the system102 and components thereof.

[129] At block 902, the method 900 includes receiving, by a receiving engine, a pluralityof feature vectors associated with data as an input.

[130] At block 904, the method 900 includes performing, by a partitioning engine, anadaptive stream partitioning of the plurality of features into a plurality of streams.

[131] At block 906, the method 900 includes processing, by a parallel processing engine,the plurality of streams parallelly to one another to generate a plurality of processed streams.

[132] At block 908, the method 900 includes performing, by a fusion engine, a sparseadaptive fusion on the plurality of processed streams to concatenate the plurality of processedstreams into a fused vector.

[133] At block 910, the method 900 includes classifying, by a classification engine, thedata based into one or more classes based on a probability assigned to each stream in the fusedvector.

[134] The present subject matter includes a number of advantages, mentioned below:a. Information-Theoretic Foundation: Grounded in Shannon entropy principles,providing the oretical justification beyond heuristics.b. Dual-Level Filtering: Filters both features (via CE) and samples (via AEE),addressing com plementary noise sources.c. Noise Robustness: Entropy-based filtering inherently handles outliers,imbalanced data, and noisy samples. Demonstrated on Adult UCI with 3:1 classimbalance.d. Edge Deployment: Computationally efficient (~15 KB model), suitable for IoTand resource constrained devices.e. Interpretability: Attention weights directly indicate sample confidence andfeature importance.f. End-to-End Learning: Fully differentiable for gradient-based optimization.g. Modularity: Components usable independently for different applications.h. Cross-Domain Generalization: Validated on two fundamentally differentdomains (biomedical: 99.20%, socioeconomic: 85.51%), proving universalapplicability.i. Parameter Efficiency: 125x fewer parameters than TabNet while maintainingcompetitive accuracy.j. Superior Classification Performance: Competitive or superior accuracycompared to existing methods (85.51% Adult UCI, 99.20% WESAD).k. Noise Robustness: Explicit entropy-based filtering for handling outliers andimbalanced data. Successfully handles 3:1 class imbalance on Adult UCI.l. Interpretability: Direct entropy-based confidence scores for model predictionsm. Flexibility: Adaptable to various data formats and domains (demonstrated ontwo fun damentally different domains).n. Scalability: Scales to high-dimensional feature spaces (validated on 14-dimensional Adult UCI and 8-12 dimensional WESAD).

[135] The present subject matter includes a number of utilities:

[136] Healthcare & Medical Diagnostics:a. ECG-based stress detection from wearable sensors (validated on WESAD:99.20%).b. Disease prediction from structured patient data.c. Anomaly detection for rare diseases.d. Outlier filtering for data entry error removal.

[137] Financial Servicesa. Credit risk assessment and loan default prediction (target application for AdultUCI validation).b. Fraud detection in transactionsc. Stock price movement forecastingd. Portfolio risk analysis

[138] Industrial IoT & Predictive Maintenancea. Sensor anomaly detection.b. Remaining useful life (RUL) predictionc. Equipment failure predictiond. Condition monitoring on edge devices

[139] E-commerce & Recommendationsa. User behavior classificationb. Personalized product recommendationsc. Purchase fraud detectiond. Customer churn predictionEXPERIMENTAL VALIDATION WITH RESULTS:

[140] Multi-Dataset Validation: The system has been validated on two independentdatasets from fundamentally different domains, demonstrating universal applicabilityand robustness:a. Dataset 1: WESAD (Wearable Stress and Affect Detection)b. Dataset2: Adult UCI (Socio economic Income Prediction)c. Cross-Domain Generalization Analysis

[141] Interpretation: The consistent high performance across fundamentally differentproblem domains (biomedical detection vs. socio economic prediction, continuous vs.mixed features, balanced vs. imbalanced classes) validates that the system is auniversal, domain-agnostic architecture rather than an approach over fitted to specificproblem types.

[142] Imbalanced Data Robustness: Adult UCI data set exhibits significant classimbalance (75.9% majority, 24.1% minority). The system's performance demonstratessuperior robustness:

[143] Majority Class Recall (94.42%): High sensitivity to majority class preventsunderfitting.

[144] Minority Class Recall (57.40%): Non-trivial performance on minority classdespite imbalance.

[145] Balanced AUC-ROC (0.9086): Superior discrimination across classificationthresholds.

[146] Entropy-Based Filtering: AEE layer explicitly identifies uncertain samples,enabling better handling of imbalanced data.

[147] Training Convergence and Stability:

[148] Observations:a. Rapid convergence: Best validation accuracy achieved by epoch 18b. Minimal overfitting: Training-validation gap only 0.23% at best epochc. Stable learning: Clean validation curve with no oscillationsd. Learning rate scheduling: ReduceLROnPlateau successfully reduced learningrate 4 times.

[149] Comparative Performance with Baselines: The system achieves competitiveaccuracy with superior parameter efficiency compared to existing methods on the AdultUCI dataset:Comparative performance on Adult UCI dataset. Uma-Attention achieves competitiveaccuracy with 125x-250x fewer parameters than TabNet and FT-Transformer,enabling edge deployment. Baseline accuracies cited from peer-reviewed benchmarks.

[150] Key Observations:a. Accuracy (Verified from Test Set): Uma-Attention's 85.51% test accuracy fallswithin the competitive range (85-86%) of gradient boosting methods andTabNet, with a maximal gap of only 1.49% from best-in-class (87% for FTTransformer).b. Parameter Efficiency(Verified): With only 3,926 trainable parameters, UmaAttention achieves:c. Edge Deployment Capability: Ultra-compact model size enables deployment onresource constrained edge devices (microcontrollers, wearables), where TabNetand FT-Transformer are impractical.d. Accuracy vs. Efficiency Trade-off: Uma-Attention achieves competitiveaccuracy within stan dard benchmarks while delivering 125x parameterreduction-a favorable trade-off for edge appli cations where memory andcomputational constraints dominate.

[151] Verified Metrics from Adult UCI Implementation:

[152] Although embodiments for the present subject matter have been described inlanguage specific to package features, it is to be understood that the present subjectmatter is not necessarily limited to the specific features described. Rather, the specificfeatures and methods are disclosed as embodiments for the present subject matter.Numerous modifications and adaptations of the system / device of the present inventionwill be apparent to those skilled in the art, and thus it is intended by the appended claimsto cover all such modifications and adaptations which fall within the scope of thepresent subject matter.

Claims

1. A method (900) for classifying data into one or more classes, comprising: receiving, by a partitioning engine (212), a plurality of feature vectors associated with data as an input; performing, by the partitioning engine (212), an adaptive stream partitioning of the plurality of feature vectors into a plurality of streams; processing, by a parallel processing engine (214), the plurality of streams parallelly to one another to generate a plurality of processed streams; performing, by a fusion engine (216), a sparse adaptive fusion on the plurality of processed streams to concatenate the plurality of processed streams into a fused vector; and classifying, by a classification engine (218), the data based into the one or more classes based on a probability assigned to each stream in the fused vector.

2. The method (900) as claimed in claim 1, comprising: receiving, by a receiving engine (208), the data as an input; and pre-processing, by a pre-processing engine (210), the data to generate the plurality of feature vectors.

3. The method (900) as claimed in claim 1, wherein the data is one of image data, video data, audio data, textual data, tabular data, sequential data, and a combination of at least two of the image data, the video data, the audio data, the textual data, the tabular data, and the sequential data.

4. The method (900) as claimed in claim 1, wherein performing the adaptive stream partitioning comprises: generating a probability matrix associated with the plurality of feature vectors via a softmax routing, wherein the probability matrix comprises a plurality of probabilities; and processing the plurality of feature vectors with each probability in the probability to generate the plurality of streams.

5. The method (900) as claimed in claim 1, wherein processing the plurality of streams parallelly comprises: performing a normalization of each streams amongst the plurality of streams to stabilize a feature magnitude associated with each stream; processing the plurality of streams having a stabilized feature magnitude via a ReLU activation; augmenting each stream with correlation energy statistics and applying a learned gating on each stream; computing a normalized Shannon entropy associated with each stream to generate the plurality of processed streams; calculating an attention weight associated with each stream based on the normalized Shannon entropy; and scaling the plurality of streams with respect to the attention weight to generate the plurality of processed streams.

6. The method (900) as claimed in claim 4, wherein computing the normalized Shannon entropy is based on: computing a Shannon entropy associated with each stream; and normalizing the Shannon entropy with a maximum Shannon entropy to generate the normalized Shannon entropy, wherein normalizing the Shannon entropy is performed to reduce noise from the plurality of streams.

7. The method (900) as claimed in claim 1, wherein performing the adaptive sparse fusion comprises: concatenating the plurality of streams with one another to generate a vector; removing a noise from the fused vector by performing a dimensionality reduction on the vector, wherein a trust score is assigned to each stream in the fused vector upon the noise removal; fusing the vector with each trust score associated with plurality of streams to generate the fused vector.

8. The method (900) as claimed in claim 1, wherein the classification comprises: performing a batch normalization of the fused vector to normalize a plurality of activations across a batch dimension; randomly dropping a predetermined percentage of the plurality of activation; and generating a probability associated with the data belonging to the one or more classes.

9. A system (102) for classifying data into one or more classes, comprising: receiving, by a receiving engine (208), a plurality of feature vectors associated with data as an input; performing, by a partitioning engine (212), an adaptive stream partitioning of the plurality of features into a plurality of streams; processing, by a parallel processing engine (214), the plurality of streams parallelly to one another to generate a plurality of processed streams; performing, by a fusion engine (216), a sparse adaptive fusion on the plurality of processed streams to concatenate the plurality of processed streams into a fused vector; and classifying, by a classification engine (218), the data based into the one or more classes based on a probability assigned to each stream in the fused vector.