Hierarchical hyperdimensional computing for energy efficient classification

Imani, Mohsen; Huang, Chenyu; Kong, Deqian; Rosing, Tajana

doi:10.1145/3195970.3196060

Cited by 33 publications

(26 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For class hypervectors with binarized elements, Hamming distance is a inexpensive and suitable similarity metric, while class hypervectors with non-binarized elements need to use cosine for similarity check. Most existing HD computing techniques are using binarized class hypervectors in order to eliminate the costly cosine metric [17,24]. However, we observed that HD with binary model provides significantly low classification accuracy as compared to non-binary model.…”

Section: Associative Memory Modulementioning

confidence: 80%

“…HD computing builds upon a well-defined set of operations with random HD vectors and it is extremely robust in the presence of failures. HD offers a complete computational paradigm that is easily applied to learning problems including: analogy-based reasoning [11], latent semantic analysis [12], language recognition [13,14], prediction from multimodal sensor fusion [15], speech recognition [16,17], activity recognition [18], DNA sequencing [19], and clustering [20].…”

mentioning

confidence: 99%

See 1 more Smart Citation

Fach

Imani

Salamat

Gupta

et al. 2019

Proceedings of the 24th Asia and South Pacific Design Automation Conference

Self Cite

View full text Add to dashboard Cite

Brain-inspired hyperdimensional (HD) computing explores computing with hypervectors for the emulation of cognition as an alternative to computing with numbers. In HD, input symbols are mapped to a hypervector and an assoc iative search is performed for reasoning and classification. An associative memory, which finds the closest match between a set of learned hypervectors and a query hypervector, uses simple Hamming distance metric for similarity check. However, we observe that, in order to provide acceptable classification accuracy HD needs to store non-binarized model in associative memory and uses costly similarity metrics such as cosine to perform a reasoning task. This makes the HD computationally expensive when it is used for realistic classification problems. In this paper, we propose a FPGA-based acceleration of HD (FACH) which significantly improves the computation efficiency by removing majority of multiplications during the reasoning task. FACH identifies representative values in each class hypervector using clustering algorithm. Then, it creates a new HD model with hardware-friendly operations, and accordingly propose an FPGA-based implementation to accelerate such tasks. Our evaluations on several classification problems show that FACH can provide 5.9× energy efficiency improvement and 5.1× speedup as compared to baseline FPGA-based implementation, while ensuring the same quality of classification. CCS CONCEPTS • Computing methodologies → Machine learning approaches; Supervised learning;

show abstract

Section: Associative Memory Modulementioning

confidence: 80%

mentioning

confidence: 99%

Fach

Imani

Salamat

Gupta

et al. 2019

Proceedings of the 24th Asia and South Pacific Design Automation Conference

Self Cite

View full text Add to dashboard Cite

show abstract

“…Thus, the rotated HD-level vector must be summed element-wise to the HD representation of the feature. This procedure has been previously described in [27,28] with the name Ngram-based encoder. It consists of differentiating the feature positions through the assignment of unique hypervectors that result from the permutation of the feature HD representations, as described by Equation (2).…”

Section: Encoding Data and Building The Classification Modelmentioning

confidence: 99%

“…For more details, we point the reader to [26], where the authors perform a comprehensive review of classification techniques based on HD computing. Indeed, HD computing was successfully applied in previous works on a limited set of research areas, like speech recognition [27,28], the internet of things [29][30][31][32], and two life-science related applications. These last attempts concern the detection of epileptogenic regions of the human brain from Intracranial Electroencephalography (iEEG) recordings [33] and pattern matching problems on DNA sequences for diagnostic purposes [34,35].…”

Section: Introductionmentioning

confidence: 99%

A Brain-Inspired Hyperdimensional Computing Approach for Classifying Massive DNA Methylation Data of Cancer

2020

View full text Add to dashboard Cite

The recent advancements in cancer genomics have put under the spotlight DNA methylation, a genetic modification that regulates the functioning of the genome and whose modifications have an important role in tumorigenesis and tumor-suppression. Because of the high dimensionality and the enormous amount of genomic data that are produced through the last advancements in Next Generation Sequencing, it is very challenging to effectively make use of DNA methylation data in diagnostics applications, e.g., in the identification of healthy vs diseased samples. Additionally, state-of-the-art techniques are not fast enough to rapidly produce reliable results or efficient in managing those massive amounts of data. For this reason, we propose HD-classifier, an in-memory cognitive-based hyperdimensional (HD) supervised machine learning algorithm for the classification of tumor vs non tumor samples through the analysis of their DNA Methylation data. The approach takes inspiration from how the human brain is able to remember and distinguish simple and complex concepts by adopting hypervectors and no single numerical values. Exactly as the brain works, this allows for encoding complex patterns, which makes the whole architecture robust to failures and mistakes also with noisy data. We design and develop an algorithm and a software tool that is able to perform supervised classification with the HD approach. We conduct experiments on three DNA methylation datasets of different types of cancer in order to prove the validity of our algorithm, i.e., Breast Invasive Carcinoma (BRCA), Kidney renal papillary cell carcinoma (KIRP), and Thyroid carcinoma (THCA). We obtain outstanding results in terms of accuracy and computational time with a low amount of computational resources. Furthermore, we validate our approach by comparing it (i) to BIGBIOCL, a software based on Random Forest for classifying big omics datasets in distributed computing environments, (ii) to Support Vector Machine (SVM), and (iii) to Decision Tree state-of-the-art classification methods. Finally, we freely release both the datasets and the software on GitHub.

show abstract

“…Imani et al [9] introduced MHD, which is a multi-encoder hierarchical classifier, facilitating HD to make the best use of multiple encoders with no increase in the classification cost. MHD comprises of two HD phase: a main phase and a decider phase.…”

Section: Literature Reviewmentioning

confidence: 99%

Various Classifiers Performance Based Machine Learning Methods

2019

IJRTE

View full text Add to dashboard Cite

Classification is a form of data mining (regarding machine learning) approach that is helpful in the prediction of group membership for data instances, where the data input is used by the computer program for learning and thereafter this learning is used for classifying the fresh observation made. This data set might just be bi-class or it can be multi-class also. Few instances of the problems in classification include: speech identification, handwriting identification, bio metric detection, document classification etc. Many classification methods exist, which can be utilized for classification. In this research work, the fundamental classification approaches and few important kinds of classification approaches that include decision tree induction, Bayesian networks,k-nearest neighbor classifier and Support Vector Machines (SVM) and fuzzy learning classifiers with their merits, drawbacks, probable applications and challenges faced with the solution available. There are different problems that have an effect on the classification and prediction. The objective of this research work is to render an extensive review of various classification approaches in machine learning. At last, the future work intended on the best classification techniques for the input data are discussed.

show abstract

Hierarchical hyperdimensional computing for energy efficient classification

Cited by 33 publications

References 14 publications

Fach

Fach

A Brain-Inspired Hyperdimensional Computing Approach for Classifying Massive DNA Methylation Data of Cancer

Various Classifiers Performance Based Machine Learning Methods

Contact Info

Product

Resources

About