MétaCan
Menu
Back to cohort
Record W7051682040

Optimization of neural networks by reducing floating- point precision of weights during training

2021· dissertation· en· W7051682040 on OpenAlexfundno aff

Bibliographic record

VenueeScholarship@McGill (McGill) · 2021
Typedissertation
Languageen
FieldEngineering
TopicElectrostatic Discharge in Electronics
Canadian institutionsnot available
FundersNatural Sciences and Engineering Research Council of Canada
KeywordsArtificial neural networkRange (aeronautics)Process (computing)Focus (optics)Table (database)Point (geometry)Representation (politics)Accuracy and precision
DOInot available

Abstract

fetched live from OpenAlex

Deep learning is an iterative process with many tunable parameters.Repeated and long running executions benefit from optimizations that improve learning and testing performance, both in terms of time and power consumption.A popular approach to optimization focuses on faster learning by clipping weight values to a range (quantization) and converting them to fixed point precision.Quantization, however, results in significant loss of accuracy and requires further optimization to reach performance of full precision models, or use other resources, such as a lookup table [23] to compensate for the information loss.In this work we focus on reducing precision more within the spirit of the native IEEE 754 standard for the representation of float values.We propose a hierarchy of approaches that reduce the number of mantissa bits of a neural network weights based on the actual values at the end of every batch of training, as simulated through a manual mantissa reduction.At a coarse level we experiment with this approach applied to an entire learning network, extending the approach to a level-by-level precision adjustment to better adapt to the internal precision needs of different layers.We further propose a novel fine-grain approach, applying precision adjustment to groups of individual weights, dynamically clustered into "buckets," each having different precision.The approach performs almost as good as a full precision network in terms of training accuracy, but allowing for reduced space requirements, and a significantly lower prediction time.i Abrégé L'apprentissage en profondeur est un processus itératif avec de nombreux paramètres réglables.Les exécutions répétées et de longue durée bénéficient d'optimisations qui améliorent les performances d'apprentissage et de test, à la fois en termes de temps et de consommation d'énergie.Une approche populaire de l'optimisation se concentre sur un apprentissage plus rapide en découpant les valeurs de poids dans une plage (quantification) et en les convertissant en précision en virgule fixe.Cependant, la quantification entraîne une perte de précision significative et nécessite une optimisation supplémentaire pour atteindre les performances des modèles de précision totale, ou utiliser d'autres ressources, telles qu'une table de recherche [23] pour compenser la perte d'informations.Dans ce travail, nous nous concentrons sur la réduction de la précision davantage dans l'esprit de la norme native IEEE 754 pour la représentation des valeurs flottantes.Nous proposons une hiérarchie d'approches qui réduisent le nombre de bits de mantisse d'un poids de réseau neuronal basé sur les valeurs réelles à la fin de chaque lot d'entraînement, comme simulé par une réduction manuelle de la mantisse.À un niveau grossier, nous expérimentons cette approche appliquée à tout un réseau d'apprentissage, en étendant l'approche à un ajustement de précision niveau par niveau pour mieux s'adapter aux besoins de précision internes des différentes couches.Nous proposons en outre une nouvelle approche à grain fin, appliquant un ajustement de précision à des groupes de poids individuels, regroupés dynamiquement en " seaux ", chacun ayant une précision différente.L'approche fonctionne presque aussi bien qu'un réseau de précision complète en termes de précision d'entraînement, mais permet des besoins d'espace réduits et un temps de prédiction nettement plus court.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.

How this classification was reachedexpand

Full frame machine prediction

Teacher imitation

Not calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.

metaresearch head score (Codex)0.001
metaresearch head score (Gemma)0.004
Version: metacan-v3-hybrid-931329e0061cValidation status: machine_predicted_unvalidated
Candidate categoriesnone
Consensus categoriesnone
DomainCandidate signal: none · Consensus signal: none
Study designCandidate signal: Simulation or modeling · Consensus signal: Simulation or modeling
GenreCandidate signal: Methods · Consensus signal: Methods
Teacher disagreement score0.010
Threshold uncertainty score0.019

Distilled classifier scores by category (both heads)

CategoryCodexGemma
Metaresearch0.0010.004
Meta-epidemiology (narrow)0.0010.001
Meta-epidemiology (broad)0.0010.000
Bibliometrics0.0000.000
Science and technology studies0.0000.000
Scholarly communication0.0010.001
Open science0.0010.001
Research integrity0.0010.001
Insufficient payload (model declined to judge)0.0040.001

Machine scores (provisional)

The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.

Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.

Opus teacher head0.007
GPT teacher head0.211
Teacher spread0.203 · how far apart the two teachers sit on this one work
Validation statusscore_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it

Classification

machine, unvalidated

Machine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.

The models applied no category: nothing in the taxonomy fit this work.
Study designSimulation or modeling
Domainnot available
GenreMethods

How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".

Quick stats

Citations0
Published2021
Admission routes1
Has abstractyes

Explore more

Same venueeScholarship@McGill (McGill)Same topicElectrostatic Discharge in ElectronicsFrench-language works237,207