Fakultät Gestaltung
Filtern
Erscheinungsjahr
- 2021 (2)
Dokumenttyp
Sprache
- Englisch (2)
Volltext vorhanden
- ja (2)
Schlagworte
Institut
Abstract. This study compares performance of different speech com-
mand classification systems which can be executed on an Raspberry
Pi Zero ARMv6 architecture. Three systems are evaluated: first one,
TREELITE_MFCC, is based on Treelite system cross-compiles a MFCC-
classifying Random Forest into a highly optimized shared library; second,
TFLITE_MFCC, performs classification of MFCC inputs by means of
convolutional neural network encoded as a lightweight TensorFlow Lite
model while the third one - labeled as TFLITE_RAW - uses more com-
plex network to directly classify the audio signal. We evaluate models
not only in terms of their accuracy and precision, but also in terms of
execution time, voltage, current and energy consumed. We observe that
while TFLITE_RAW offers superior performance in terms of accuracy,
TREELITE_RAW is also worth of consideration for low-latency real-
life applications since it offers relatively good performance (avg. micro-
precision=0.917) but its predictions, when executed on Raspberry Pi
Zero, are significantly faster and cost less energy than TensorFlow Lite
models.
The paper presents a novel method of multiclass classification. The method combines the notions of dimensionality reduction and binarization with notions of category prototype and evolutionary optimization. It introduces a supervised machine learning algorithm which first projects documents of the training corpus into low-dimensional binary space and subsequently uses canonical genetic algorithm in order to find a constellation of prototypes with highest classificatory pertinence. Fitness function is based on a cognitively plausible notion that a good prototype of a category C should be as close as possible to members of C and as far as possible to members associated to other categories. In case of classification of documents contained in a 20-newsgroup corpus into 20 classes, our algorithm seems to yield better results than a comparable deep learning "semantic hashing" method which also projects the semantic data into 128-dimensional binary (i.e. 16-byte) vector space.