Bio-inspired Convolutional Neural Networks

Researching the new generation of Deep Neural Networks for applied Computer Vision

About BioCNN

The Bio-inspired Convolutional Neural Network is an experimental deep learning architecture that has been introduced in 2017, but has longer history with its inception dating back in 2010. The research investigates if this method:

  • Could reduce memory requirements for CNN models.
  • Increase generalization capabilities by reducing overfitting during training.
  • Has sufficient performance for real-world applications in the Internet of Things (IoT)

Moreover, experiments followed by formal justifications, could discover new Computational Intelligence methods that are inspired by and based on knowledge from Visual Neuroscience.

Research

History and Present

I have started working on bio-inspired neural networks during my BSc thesis "Visual Objects Recognition using Object-oriented Neural Networks". Convolutional Neural Networks, which are inherently bio-inspired, were used for my Master's thesis "Content-based Image Retrieval Using Deep Learning", that is available on ResearchGate. Now I am expanding my knowledge and understanding by doing focused research as a PhD candidate in the Artificial Intelligence and Information Analysis (AIIA) Lab at the Artistotle University of Thessaloniki.

Deep Learning for Computer Vision

Convolutional Neural Networks have showcased their superiority as learnable visual feature extractors in the ImageNet Large Scale Visual Recognition Challenge. Nevertheless there are several open research issues, like understanding the proper choice of architectural hyperparameters that will create a lightweight model with sufficient performance on a specific Computer Vision problem.

Web Intelligence

My domain of expertise combines the World Wide Web with Artificial Intelligence, in the next generation that could be described as Web 4.0 (Web 3.0 is used to describe the Semantic Web). Machine Learning is in the core of Web Intelligence, but in order to be used in innovations it is combined with Web Software Engineering and User Experience Design. Additionally Mobile Networks and Mobile/Embedded Software Development is needed for future solutions, realizing the concept of Ubiquitous (Pervasive) Computing.

Machine Learning Tools

Computational Frameworks for Accelerated Computing

Tensorflow is a free computational framework developed by Google that could be used for Machine Learning and other parallelized computations. Caffe from Berkeley AI Research (BAIR) has a plethora of Convolutional Neural Network implementations that can be found in its model zoo. Both of them can be used in a CUDA enabled GPU to accelerate Deep Learning.

Tensor Abstraction Layer Objects - TALOS

TALOS is a new deep learning library in Python that makes it easy for a begginer to create and train neural networks. It is a software abstraction layer on top of Tensorflow, sharing the same principles with Keras. Moreover, it has a queue system to execute non-stop training sessions and a multi-threaded design for loading pages of samples from a disk chache into the main-memory. Furthermore, there is the concept of "experiment" that includes various information like, stats during training, evaluation of the trained model, generation of graphs and many more. I have started its development at the time I was writting my MSc Thesis and its alpha version will be shortly available on GitHub.

Datasets

You can find links, informations and usage instructions for Computer Vision datasets here, plus all my publicly available datasets including LITE, that I am using to train BioCNNs. You can download the datasets and start training them in TALOS, for your "Hello World" in the domain of Deep Learning for Computer Vision

Contact me

Feel free to contact me at my email