Groundbreaking research enabling applied AI solutions. conducts ground-breaking yet practical AI Research so you can build your AI Solution rapidly.

AI Research

Our mission is to build practical and groundbreaking AI Research so companies and professionals can apply to their production systems. Our goal is to provide AI Solutions by implementing our algorithms, tools and technologies.

Our researchers and engineers are dedicated to working towards this goal and they contributes relentlessly towards building practical software and algorithms. We publish our research and present at leading conferences regularly but our differentiation is in applying it directly into production systems of industry verticals such as healthcare diagnostics, manufacturing operations and more.


Our mission is to build practical and groundbreaking AI Research which companies and professionals can directly apply to their production systems. Our goal is to provide AI Solutions by implemnting our algorithms, tools and technologies.

Our researchers and engineers are dedicated to working towards this goal. To do so our team contributes relentlessly towards building practical software and algorithms.

We publish our research and present at leading conferences regularly but
adopt a unique strategy by applying it directly into industry verticals.
We believe that is the only way to walk the talk!

3rd October 2020

An energy efficient time-mode digit classification neural network implementation

This paper presents the design of an ultra-low energy neural network that uses time-mode signal processing). Handwritten digit classification using a single-layer artificial neural network (ANN) with a Softmin-based activation function is described as an implementation example. To realize time-mode operation, the presented design makes use of monostable multivibrator-based multiplying analogue-to-time converters, fixed-width pulse generators and basic digital gates. The time-mode digit classification ANN was designed in a standard CMOS 0.18 μm IC process and operates from a supply voltage of 0.6 V. The system operates on the MNIST database of handwritten digits with quantized neuron weights and has a classification accuracy of 88%, which is typical for single-layer ANNs, while dissipating 65.74 pJ per classification with a speed of 2.37 k classifications per second. This article is part of the theme issue ‘Harmonizing energy-autonomous computing and intelligence’.

Link to Paper , Learning path and Github Code 

1st October 2020

Project DentAI

Artificial intelligence for dental image analysis: A Guide for Authors and Reviewers

Objectives: The number of studies employing artificial intelligence (AI), specifically machine and deep learning, for dental image analysis is growing fast. The majority of studies suffer from limitations in planning, conduct and reporting, resulting in low robustness and applicability. We here present a non-authorative guide for authors and reviewers to be applied, discussed and further developed.

Methods: Lending from existing reviews in other fields and founded on the principles of evidence-based research practice, a set of guidance items are presented, assisting future scientists, reviewers and editors in planning, conducting, reporting and evaluating studies on AI in dental image analysis. The items have been derived on a discussion basis within the ITU/WHO focus group “Artificial Intelligence for Health (AI4H)”, and the topic group “Dental diagnostics and digital dentistry” and should be rigorously appraised and adapted.

Results: Thirty-one items on planning, conducting and reporting studies were devised. These involve items on the study’s wider goal, focus, design and specific aims, data sampling and reporting, sample estimation, reference test construction, model parameters, training and evaluation, uncertainty and explainability, performance metrics and data partitions.

Conclusion: Scientists, reviewers and editors should consider this guide when planning, conducting, reporting and evaluating studies on AI for dental image analysis.

Clinical significance: Current studies on AI in dental image analysis show considerable weaknesses, hampering their replication and application. This non-authorative guide may assist scientists, reviewers and editors to overcome this issue and advance AI research in dentistry as well as facilitate a forward-debate on standards in this fields.

Authors: Falk Schwendicke1,2, Tarry Singh2,3, Jae-Hong Lee2,4, Robert Gaudin5, Joachim Krois1,2

Link to published (to be updated asap) Paper, Learning Path, and Github Code(once it is released).

September 21st, 2020

Seismic Facies Analysis using state of the art architecture: A deep domain adaptation approach

Deep neural networks (DNNs) are powerful tools that are able to learn accurately from large quantities of labeled input data. However DNNs cannot always generalize on test data sampled from different input distributions. We demonstrate the use of unsupervised Deep Domain Adaptation (DDA) in a DNN when no input labels are available and distribution shifts are observed in target domain (TD). Experiments are performed on seismic images of F3 block 3D dataset from offshore Netherlands as source domain (SD) and Penobscot 3D survey data from Canada as TD. Three geological classes fromSD and TD that have similar depositional environment and lithology are considered for the study. To approach the studywe developed a deep neural network architecture, EarthAdaptNet(EAN), specially designed for semantically segmenting the seismic images with a minimal amount of training data. More specifically, we use a transposed residual unit to replace the traditional dilated convolution in the decoder block. EarthAdaptNet shows promising results in comparison to baseline results and was able to achieve a pixel accuracy>84% and an accuracy of70% for the minority classes, an improvement on the pre-existing architectures.

Presented at Geoconvention Conference, Canada Sept 21 – 23 2020

Read Detailed Blog Article, Link to Paper , Learning path and Github Code (once it is released).

March 30th, 2020

Project CytoDeep

Feasibility Assessment of Artificial Intelligence in Breast Cancer Diagnostics

The issue of breast cancer is one which affects millions around the globe each year with those in developing regions facing disproportionately high rates of fatality. It is known the importance of early diagnosis, but with the relative scarcity and high workload of pathologists in these developing regions, it is vital to research how recent advances in artificial intelligence can serve to aid in the diagnosis of breast cancer. It has been shown that deep learning models already have the capacity for rapid diagnosis with many having the same performance as health-care professionals. In the field of breast cancer, there have been a variety of deep learning diagnostic models created; however, there are very few which have been made to diagnose FNAC results. Since breast cancer diagnosis through FNAC is well suited for developing countries due to its minimal cost and infrastructure requirements, developing a deep learning diagnostic model could be an invaluable tool to assist pathologists working in these regions.

There are a number of challenges associated with building the proposed diagnostic model, but the primary issue is the need for a large, balanced and properly labelled dataset. In the Materials and Methods section, a variety of techniques were discussed in detail to deal with these issues commonly faced with image classification tasks. The majority of these procedures such as regularization to prevent model overfitting, data augmentation to increase diversity of training examples, imputation of missing data entries, oversampling to correct for class imbalance, cost-sensitive learning to minimize false negative diagnosis, semi-supervised learning to handle images with missing labels, and transfer learning are fairly common procedures which many deep learning medical diagnostic models have employed to increase performance.

We will soon share the link to Detailed Blog Article, Link to Paper and Github Code (once it is released).

May 24th, 2019

Towards a Neurobiological Basis of Deep Learning

The human brain constantly executes myriad decisions every day – some trivial, many complex. Decision making and learning are fundamental to human survival, and our runaway success as a dominant species. Our brain continually decides by reflecting upon past experiences, while simultaneously acquiring new knowledge with every decision. Neuromodulators – acetylcholine (ACh), noradrenaline (NA), serotonin (5-HT), dopamine (DA), and histamine (HA) – reorganize the function of local neural networks neurons and shape the emergence of global brain states such as decision making and learning. The advent of artificial neural networks, which has benefited from the remarkable success of the brain’s ability to decide and learn, is attempting to transform human society through machine-based representations that mimic patterns of biological neural activity. For example, biologically inspired convolutional neural networks (CNNs) have shown promising performance in a variety of tasks including image recognition, classification and analysis. Recent studies have adopted a more biologically-realistic compartmental structure in the design of deep learning algorithms. Here, we review subcortical structures and neuromodulatory systems that regulate contextual decision making and learning in the brain, and outline proposals towards more efficient machine-based representations for neuromodulation-aware models of deep-learning. Taken together, a comprehensive review of existing findings on the role of neuromodulators in decision making and learning processes will be essential for evidence-driven, biologically-inspired deep learning models.

May 8, 2019


Generative Model with Dynamic Linear Flow

Flow-based generative models are a family of exact log-likelihood models with tractable sampling and latent-variable inference, hence conceptually attractive for modelling complex distributions. However, flow-based models are limited by density estimation performance issues as compared to state-of-the-art autoregressive models. Autoregressive models, which also belong to the family of likelihood-based methods, however suffer from limited parallelizability. In this paper, we propose Dynamic Linear Flow (DLF), a new family of invertible transformations with partially autoregressive structure. Our method benefits from the efficient computation of flow-based methods and high density estimation performance of autoregressive methods. We demonstrate that the proposed DLF yields state-of-the-art performance on ImageNet 32×32 and 64×64 out of all flow-based methods,and is competitive with the best autoregressive model. Additionally, our model converges 10 times faster than Glow (Kingma and Dhariwal, 2018).

Links to Detailed Blog Article, Link to Paper and Github Code

May 27, 2018

Intra-thalamic and Thalamocortical Connectivity: Potential Implication for Deep Learning

Contrary to the traditional view that the thalamus acts as a passive relay station of sensory information to the cortex, a number of ex-perimental studies have demonstrated the effects of peri-geniculate and cortico-thalamic projections on the transmission of visual in- put. In the present study, we implemented a mechanistic model to facilitate the understanding of perigeniculate and corticothalamic effects on the transfer function of geniculate cells and their firing patterns. As a result, the model successfully captures some funda- mental properties of early-stage visual processing in mammalian brain. We conclude, therefore, that the thalamus is not a passive relay center and the intra-thalamic circuitry is of great importance to biological vision. In summary, intra-thalamic and thalamocortical circuitry has implications in early-stage visual processing, and could constitute a valid tool for refining information relay and compression in artificial neural networks (ANN), leading to deep learning models of higher performance.
Coming up
May 22, 2018


Utilizing Richard’s Curve for Controlling the Non-monotonicity of the Activation Function in Deep Neural Nets

This work introduces a novel activation unit that can be efficiently employed in deep neural nets (DNNs) and performs significantly better than the traditional Rectified Linear Units (ReLU). The function developed is a two parameter version of the specialized Richard’s Curve and we call it Adaptive Richard’s Curve weighted Activation (ARiA). This function is non-monotonous, analogous to the newly introduced Swish, however allows a precise control over its non-monotonous convexity by varying the hyper-parameters. We first demonstrate the mathematical significance of the two parameter ARiA followed by its application to benchmark problems such as MNIST, CIFAR-10 and CIFAR-100, where we compare the performance with ReLU and Swish units. Our results illustrate a significantly superior performance on all these datasets, making ARiA a potential replacement for ReLU and other activations in DNNs.

Coming soon: DeepSwitch

The solutions found by the adaptive algorithms like Adam fail to generalize as well as SGD in certain scenarios even though adaptive methods usually perform well on training set. So, there is often tradeoff between testing accuracy and performance update at local optimas. Keskar et. al have showed that the adaptive methods work better in the initial portion of the training but with later portion, SGD seems to work better. The basic premise of this work is to investigate the use of fuzzy logic to extend the phenomenon to a more generic, and robust control for optimizer switching. Unlike the prior work, we also incorporate quasi-Newtonian optimizers as well as other adaptive optimizers than Adam, and work out a switching logic that maximizes the generalization accuracy while having minimal effect on training time.
Coming up

Join us

Our DeepRRP (Deep Learning Remote Residency) Program is currently in beta open for AI Researchers who are already conducting some form of research in Machine Learning, Deep Learning or Artificial Intelligence.

For more information on our selection criteria, please click the enroll button below.