NeuTigers Inc’s edge-AI algorithms bring AI to resource-constrained devices. This makes AI applications more personalized, scalable and in real time with a higher level of privacy and security than conventional AI. The company is co-founded by Princeton Engineering research.

NeuTigers’ proprietary IP overcomes the traditional challenges of ML that require large datasets,

Expensive compute power and high latency with inference on the cloud.

Its breakthrough approach “shrinks” Neural Network models to be deployed directly on mobile and wearable devices rather than in the cloud. It also reduces the need for large datasets with its innovative approach to high accuracy synthetic data generation.

A Look at NeuTigers IP

Making neural networks smaller and efficient (NeST, ChamNet, and TUTOR frameworks):

  • Randomly initialized sparse network (seed architecture)
  • Iterative tuning of the architecture with gradient-based growth and magnitude-based pruning of neurons and connections

NeST yields accurate yet very compact DNNs, with a wide range of seed architecture selections.

  • Reduce network parameters by 70.2× (74.3×) and floating-point operations (FLOPs) by 79.4× (43.7×) for LeNet-300-100 (LeNet-5) architecture
  • Reduce network parameters (FLOPs) by 15.7× (4.6×) and 30.2× (8.6×), for the AlexNet and VGG-16 architectures
  • NeST’s grow-and-prune paradigm delivers additional parameters and FLOPs reduction relative to pruning-only methods.

The ChamNet framework provides an efficient, scalable, and automated Neural Network (NN) architecture adaptation methodology. This adaptation reduces search time. It employs predictive models (accuracy, latency, and energy predictors) to speed up the entire search process by enabling direct performance metric estimation.

Refer to the following articles for details and performance analysis: NeST, ChamNet

Synthetic data

Our synthetic data generation toolkit relies on the TUTOR Deep Neural Networks (DNNs) synthesis framework. 

TUTOR produces synthetic data in three steps: 

  • Drawing synthetic data from the same probability distribution as the training data and labeling the synthetic data based on a set of rules extracted from the real dataset.
  • Two training schemes combine synthetic data and training data to learn DNN weights.
  • Employing a grow-and-prune synthesis paradigm to learn both the weights and the architecture of the DNN to reduce model size while ensuring its accuracy. 

NeuTigers synthetic data library  

Reduces the need for labeled data by 5.9X. Improves output accuracy by 3.4% despite using a smaller sample size during the learning phase.
Uses fewer samples than Generative Adversarial Networks (GANs).

Refer to the following article for details and performance analysis: TUTOR 

Read about our SynthDeep product

Synthetic Data Brochure

Download our Synthetic Data Brochure (854 KB)

Our Secret Sauce

Improve Energy Efficiency 

10 – 100 NanoJoules

We optimize battery usage time

Optimize Latency Efficiency


We reduce delays in processing & decision making, and improve customer experience

Increase Accuracy

We increase sample-sized and accuracy with synthetic data augmentation approach

Scale Portability

Cloud server to mobile and wearable

 Any hardware

Improve access to personalized services while preserving privacy & security


Most of the Cyber-Physical Systems (CPS) and Internet-of-Things (IoT) devices are energy-constrained, which makes them unable to implement elaborate cryptographic protocols and primitives and other conventional security measures across the software, hardware, and network stacks.

Another challenge in securing CPS/IoT is the large amount of accessible data generated by the numerous communication channels among devices. 

NeuTigers’s ML-based approach systematically generates new exploits in a CPS/IoT framework. We call this approach SHARKS (Smart Hacking Approaches for RisK Scanning).

ML operates at both system and user levels to predict unknown exploits against CPS/IoT. 

  • SHARKS models the behavior of CPS/IoT under attack at the system and network levels,
  • Use ML to discover a more exhaustive potential attack space
  • Map it to a defense space.  

The GRAVITAS model comes into play to secure CPS/IoT devices.

  • GRAVITAS overcomes the challenge of having limited resources available to these devices by combining a system’s hardware, software, and network stack vulnerabilities into a single attack graph. 
  • Attack graph produces probabilistic “vulnerability score” describing an attack’s appeal to different classes of adversaries.
  • Subset of user-defined defenses lowers the vulnerability score
    GRAVITAS presents a security model  

Refer to the following articles for details and performance analysis: SHARKS, GRAVITAS

Health assessment through wearables

The core IP presented above makes real-time disease detection and long-term monitoring through wearable sensors more efficient, private and secure. 

We have two platforms in this space – CareDeep and StarDeep. 

  • CareDeep: A community-focused care chronic diseases coordination platform for better patient engagement including a patient chatbot and a provider dashboard for real-time monitoring and personalized care. Read our about our SickleDeep product 
  • StarDeep: Regulated diagnostics to detect early signs of various disease symptoms before flare-ups, clinical exacerbation, and, ultimately, hospitalization.


By combining wearable medical sensors (WMSs) and machine learning, StarDeep makes it possible to detect and diagnose diseases in a non-invasive way.

This approach exploits the superior knowledge distillation capability of machine learning to extract medical insights from health-related physiological signals, establishing digital biomarkers for various disease states. To collect physiological signals, we use commercially available devices such as smartwatches, smartphones, and readings from galvanic skin response, oxygen saturation, and blood pressure sensors. After training the models, our model only requires two minutes of data collection for local inference on the device, yielding real-time results that preserve privacy. 

StarDeep models can be adapted to different models of smartwatches and smartphones. 

Each of the StarDeep products undergoes rigorous clinical trials under the supervision of trained clinicians. All the physiological signals and questionnaire data are collected under Institutional Research Board (IRB) approval. 

The suite of Healthcare products includes:


CovidDeep detects Covid-19 status using device data as explained above and some easy-to-answer questions in a questionnaire whose answers can be obtained through a smartphone application. We have reached a test accuracy of 98.1%.

CovidDeep performs well on  false positive rate, false-negative rate, and F1 score.   

Refer to the following article for details and performance analysis: CovidDeep

Read the CovidDeep case study

Mental Health: MHDeep

MHDeep diagnoses three important mental health disorders: schizoaffective, major depressive, and bipolar. This has the potential to enhance the ability of the physician to intervene quickly when mental health conditions deteriorate.  

Refer to the following article for details and performance analysis: MHDeep 


DiabDeep classifies a person’s diabetes status: type 1, type 2 or healthy. The model achieves a 96.3% (95.3%) accuracy in classifying people with diabetes against healthy individuals. 

Refer to the following article for details and performance analysis: DiabDeep

Your Smartphone Understands You  

Using Machine Learning to Address Fundamental Human Needs   

NeuTigers has developed an ML model embedded in smartphones to understand its user. We call it “Your Smartphone Understands You (YSUY).” YSUY uses wearable medical sensors to understand physical, mental, and 4-class (2-class) emotional states with 91.2%, 91.1%, and 96.9% (99.1%) accuracy.     

YSUY is a promising candidate for adapting ML models to human-centric needs. We have developed a smartphone app for YSUY and performed experiments with it in real-life situations without limiting the users to specific experimental protocols.  

Refer to the following article for details and performance analysis: YSUY