Source  
Study identifies a universal property for efficient communication On April 16, Friday, 2021

A study published on 23rd March in the journal Proceedings of the National Academy of Sciences analyzed how artificial neural networks develop spontaneous systems to name colors. The authors found that the emerging color vocabulary has exactly the same property of optimizing the complexity/accuracy trade-off found in human languages. This suggests that an efficient categorization of colors (and possibly other semantic domains) in natural languages is not dependent on specific human biological constraints.

Read more at www.ethicaleditor.com

Source  
CPU algorithm trains 15 times faster than top GPU trainers On April 08, Thursday, 2021

Rice University computer scientists develop software that runs on commodity processors. Their software trains deep neural networks 15 times faster than platforms based on graphics processors. The research will be presented April 8 at the machine learning systems conference MLSys. Deep neural networks are powerful forms of artificial intelligence that can outperform humans at some tasks. The cost of training is the actual bottleneck in AI, an assistant professor says, and companies are spending millions to train them. Rice's "sub-linear deep learning engine" (SLIDE) is specifically designed to run on commodity CPUs.

Read more at techxplore.com

Source  
Healthy skin to be set apart from Diseased skin - Stern step from ANN On April 07, Wednesday, 2021

A new deep neural network architecture can provide early diagnosis of systemic sclerosis. Systemic sclerosis is a rare autoimmune disease marked by hardened or fibrous skin and internal organs. Early diagnosis is critical for patients with SSc, but often elusive. The proposed network could easily be implemented in a clinical setting, providing a simple, inexpensive and accurate screening tool for SSc. The work is published in the IEEE Open Journal of Engineering in Medicine and Biology.

Read more at www.sciencedaily.com

Source  
Artificial Intelligence tool can help detect melanoma On April 05, Monday, 2021

Melanoma is a malignant tumor responsible for more than 70 percent of all skin cancer-related deaths worldwide. Early-stage melanoma identification can improve melanoma prognosis and significantly reduce treatment cost. Researchers have developed a new artificial intelligence pipeline to more quickly and efficiently identify skin lesions. The pipeline uses deep convolutional neural networks (DCNNs) to analyze skin lesions using wide-field photography common in smartphones and personal cameras. The system uses DCNNs to optimize the identification and classification of pigmented lesions.

Read more at news.mit.edu

Source  
3D fluorescence microscopy gets enhancement using Recurrent Neural Network On March 24, Wednesday, 2021

UCLA researchers develop deep learning-enabled volumetric microscopy framework for 3D imaging of fluorescent samples. New method only requires a few 2D images of the sample to be acquired for reconstructing its 3D image. Convolutional recurrent neural network that is at the heart of this 3D fluorescence imaging method intuitively mimics the human brain in processing information and storing memories. These advances offer much higher imaging speed for observing 3D specimen, while also mitigating photo-bleaching and photo toxicity related challenges.

Read more at www.biophotonics.world

Source  
New Lexical Semantic Influence Network finds innovation from 19th century newspapers On March 22, Monday, 2021

In the 1800s in the US, the movement to end slavery found a voice in abolitionist newspapers. Researchers from Google, Georgia Tech and Emory University studied these papers. They found papers that took the lead in new usages of specific words. Researchers used machine learning to identify papers that were at the vanguard of language change. The findings provide an overall picture of the network of semantic influence among the publications.

Read more at syncedreview.com

Source  
Novel technology to improve learning ability of ANN chips On March 16, Tuesday, 2021

ANN chips are capable of mimicking the structural, functional and biological features of human neural networks. Researchers from UNIST and Tsinghua University developed a new learning method to improve the learning ability of artificial neural networks by challenging its instability. The learning method developed through an experiment to classify handwriting composed of numbers 0-9 has an effect of improving learning ability by about 3%.

Read more at techxplore.com

Source  
LEAF - A Learnable Frontend for Audio Classification On March 15, Monday, 2021

Machine learning models for audio understanding have seen tremendous progress over the past several years. Deep neural networks are used to recognize speech, understand music, or classify animal vocalizations. They rely on pre-processed data in the form of mel filterbanks, designed to replicate some aspects of the human auditory response. LEAF - LEarnable Audio Frontend, is a neural network that can be initialized to approximate mel filterbanks, and then be trained jointly with any audio classifier to adapt to the task at hand.

Read more at ai.googleblog.com

Source  
COIN - Oxford's Image Compression method better than JPEG at low Bitrates On March 14, Sunday, 2021

Oxford researchers have developed an image compression approach, COIN that outperforms the JPEG standard at low bitrates. The approach stores the weights of a neural network overfitted to the image, rather than the RGB values for each pixel. It can decode large images with surprisingly small networks (8k parameters). The researchers tested their method on Kodak images of size 768 x 512 to test it against other compression methods.

Read more at syncedreview.com

Source  
Turing Patterns fool Neural Networks On March 13, Saturday, 2021

Skoltech researchers show that patterns that can cause neural networks to make mistakes in recognising images are, in effect to Turing Patterns found all over the natural world. In the future, this result can be used to design defences for pattern recognition systems currently vulnerable to attacks. The paper was presented at the 35th AAAI Conference on Artificial Intelligence (AAAI-21) at Skoltech in New York City.

Read more at www.eurekalert.org

Source  
Sparse Inference - Speeding up Mobile and Web based Neural Network On March 11, Thursday, 2021

The application of neural nets to draw on-device inferences enables a variety of real-time applications, like pose estimation and background blur. One way to optimize a model is through use of Sparse Neural Networks. There is a release of new features for the XNNPACK acceleration library and TensorFlow Lite that enable efficient inference of sparse networks. These tools power a new generation of live perception experiences, including hand tracking and background features in Google Meet.

Read more at ai.googleblog.com

Source  
Automated Feature Engineering Using Neural Network On February 28, Sunday, 2021

Automation processes are implemented to maintain time efficiency and reduce tediousness. Feature engineering, an integral part of EDA (Exploratory Data Analysis) enables the selection of interactive features in the learning model. This methodology allow the days long process of feature engineering automate while implementing learning models.

Read more at towardsdatascience.com

TAGS