LACROIX Impulse specializes in Artificial Intelligence, Machine Learning, Deep Learning and Computer Vision.
Artificial Intelligence
LACROIX Impulse specializes in Artificial Intelligence, Machine Learning, Deep Learning and Computer Vision & we offer customized support to companies willing to integrate AI based applications.
Artificial intelligence covers the domain of computer systems able to perform tasks normally requiring human intelligence or animal intelligence, such as visual perception, speech recognition, decision-making, and translation between languages.
- Natural language processing (text to speech, speech to text),
- Visual perception (machine vision, image recognition)
- Social intelligence (human feeling, emotion and mood simulation)
- Planning and optimization (intelligent agent, multi-agents)
- Expert systems (problem solving)
- Robotics (motion and manipulation)
Machine Learning & Deep Learning
Machine Learning is a subdivision of AI while Deep Learning is a subdivision of Machine Learning. Machine learning focuses on enabling systems to perform tasks without explicit programming. Deep Learning does the same, but it is based on artificial neural networks (ANN).
Impulse has been developing its skills in Machine Learning and Deep Learning to design the most relevant customer solutions whether algorithms are located on edge or on premise.
Systems based on machine learning and deep learning require a two phases approach:
- First one is called the training phase. It refers to the process of creating a machine learning algorithm/model and making it learn.
- The second one is the inference phase and refers to the process of using a trained machine learning algorithm to make a prediction.
Two complementary technologies
Machine learning and deep learning can’t be opposed to each other. Deep learning can analyze images, videos, and unstructured data in ways machine leaning can’t easily do. It requires less ongoing human intervention and offers better performance but it also requires a lot more computational power than machine learning.
Deep learning divides the car recognition task into various layers: one algorithm layer learns how to recognize the tires, another one the wheel, etc. Once connected to each other, these layers have integrated an overall capacity to recognize cars so that each time a new image is presented they can indeed recognize the correct vehicle.
Drivers of Success
Machine learning can be classified according to their drivers of success in the industry:
- Supervised learning
- Transfer learning
- Unsupervised learning
- Reinforcement learning
In supervised learning and transfer learning approaches, engineer needs to train the model with data that have been previously labelled (annotated). Creating labelled data is expensive as the task is often done manually (annotating all cars, all trucks and all pedestrians in an image for instance).
While supervised learning trains the model from scratch, transfer learning leverages on an already trained model with a generic dataset and retrains the model with a limited amount of specialized labeled data. Generic dataset such as public dataset can be used but intellectual property rights and compliancy with GDPR have to be checked prior using them.
In unsupervised learning approach, there is no need to annotate the data, which does not mean there is no need for a dataset. The algorithms discover hidden patterns or data groupings without the need for human intervention. Unsupervised learning is suitable for making predictions for instance.
In reinforcement learning approach, the machine learning model is trained to make a sequence of decisions. The agent learns to achieve a goal in an uncertain, potentially complex environment. Reinforcement learning is used in domains such as gaming for instance.
Impulse has successfully combined supervised learning, transfer learning and unsupervised learning technics to make predictions of anomaly in video streams where anomaly is characterized by the prevalence and the occurrence of certain behaviors.
Experts Tradeoffs
Whenever deep learning inference needs to take place on the edge device, tradeoffs have to be done in order to keep computational processing and latency acceptable on the one hand, while keeping precision and recall as high as possible on the other hand. Processing on the edge is particularly useful to reduce overall data bandwidth necessary.
Impulse engineers have developed deep knowledge expertise in cutting neural network layers, selecting neural networks fitting with edge platforms computing capability, training networks with specific technics such as quantization aware training, post-training quantization or network pruning.
We have also gathered experience across projects on how to reduce the amount of training dataset needed to train the neural networks while optimizing the metrics results.