Smart agents, AI and automation
Artificial Intelligence and Machine Learning have been gaining more and more fame during last decade. More powerful hardware and better algorithms allowed to achieve unbelievable results in many different fields, from image classification to speech recognition, from semantic analysis to autonomous agents. During the years, such AI powered agents have been able to outperform humans in a growing number of tasks.
These technologies make their users able to automate a large, and growing, number of operations, radically transforming businesses, increasing efficiency and cutting costs as never before.
α-MaLe™ is a modular software suite, based on most advanced frameworks, which implements top performing Machine Learning algorithms in dedicated modules to execute specific tasks and solve specific problems. It is also able to leverage parallel computing and GPU CUDA technologies, to obtain maximum computation speed. It can be deployed in private environments or on a cloud infrastructure in Software-as-a-Service solution.
Image, Video and Audio Processing
α-MaLe™ provides functionalities to perform a large number of advanced autonomous tasks on images, videos and audio files, some of the most important are detailed below. It allows the user to leverage best in class algorithms to perform all these tasks, obtaining reliable and super fast results.
Classification, Detection & Localization and Segmentation
Classification is one of the most common applications of machine learning, and one of the most successful. Computers have achieved super-human performances in this field: with error rates lower than those scored by a human, they are able to correctly classify objects almost instantly.
When dealing with Detection & Localization, the algorithm provides additional info about the input data it operates on, in particular how many target instances are present, and where in the frame they are located.
Algorithms for Semantic Segmentation are able to detect, classify, locate and isolate objects of interest from the surrounding context, identifying and grouping their constitutive elements (i.e. images pixels).
Style Transfer and Patch Harmonization
Recent research has been investigating usage of machine learning techniques in image generation. In particular, deep neural networks have shown huge potential in generating high quality images in incredible contexts, like: style retrieval and transfer, where the style of an image is applied on another one, patch harmonization, where an element is added to an image and the algorithm takes care of merging it with the context applying appropriate style and texture, and photorealistic image generation.
Similar techniques can be applied on audio and video too, generating new content leveraging realistic artificial synthesis. Most recent results in this field have reached a quality level thought unbelievable only few years ago: they have been proven able to easily fool human judgment in distinguishing original images, audio or video from synthetic ones.
α-MaLe™ features different powerful algorithms able to perform these and many other similar tasks, providing the user with tools derived from the most recent and advanced research in the field. Its flexibility allows to accommodate very specific needs and requirements, with the additional possibility to combine different modules together and take advantage of combined capabilities.
One of the most interesting directions in the field of AI/ML is the control of dynamic systems. Being able to code and effectively train an agent capable of controlling a complex dynamic system, achieving human-level or super-human performances only by collecting experience from the surrounding environment, is a key milestone in the pursuit of “full autonomy”, and is the main goal of Reinforcement Learning.
From robotic systems control (rovers, drones/UAVs, humanoids), to AI-based video games enhancing, from realistic simulation scenario animation to complex non-linear systems actuation, Reinforcement Learning research is exploring many different fields, and it is collecting historic achievements.
Gaming and Simulation
Currently, entities inside games and simulators are animated by standard, rule-based, behavior and dynamics. Adopting Reinforcement Learning techniques is the most promising way to move current state-of-the-art to the next level: it will allow to dramatically boost gaming experience, and will play a major role in simulators, providing key tools to foster realism, improving training & simulation applications.
In order to obtain high quality results in this domain, various aspects must be mastered: knowledge of state-of-the-art RL techniques and algorithms, integration of accurate and reliable simulation environments, experience with classic control techniques, deep understanding of complex non-linear systems dynamics.
α-MaLe™ technology, which implements most advanced RL solutions currently available, directly stems from our research activity, which targets various contexts and builds upon relevant experience matured over the years in dynamic systems control.
Classic control techniques are still the standard solution for robotics control, even if they are increasingly often being coupled with Machine Learning solutions, especially in the perception area. At the same time, Reinforcement Learning approaches are being studied on fairly complex robotics control problems, showing promising results, both in simulation as well as when deployed on real systems, and opening opportunities to create RL-based controllers capable of dealing with problems dynamics where current classic control fails.
Recently, deep learning has increasingly found application in providing impressive “Super resolution” capabilities. Algorithms are trained on datasets where low resolution images are coupled with high resolution ones (ground truth), allowing them to generalize how to infer optimal upscaling transformations. With a proper dataset, it is possible to obtain incredible results.
Following images are obtained with the very same model, demonstrating capability to provide notable results on different scenes and contexts. In these cases the low resolution image is upscaled to obtain an image having height and width four times the original ones, thus a sixteen times increment on the original area.
α-MaLe™ technology is able to provide state-of-the-art solutions for these applications, with the additional possibility to customize them on specific datasets and contexts, maximizing outcome quality.
By applying state of the art data analysis on business information, it is today possible to gain deep strategic business insights. Tools like complex event processing, business performance management, online analytical processing, elaborating large amounts of structured and unstructured data, allow to identify, develop and create new opportunities, helping to gain market advantage and long-term stability.
It plays a role on different levels, from operational (like product positioning or pricing) to strategic (like prioritization or long-term goals definition). Maximum effectiveness is obtained when external data (i.e. related to the market segment of interest) is combined with internal one (i.e. regarding company financial aspects). Most important applications are: performance metrics and benchmarking to understand current progress towards business goals, business analytics for business knowledge discovery, and business reporting to inform and support business strategy. α-MaLe™ technology supports a large number of business intelligence applications, with the additional advantage of being fully customizable on specific user needs.
Embedded, Cloud and Desktop Solutions
Every project has its own constraints in terms of budget, computational power, execution time, physical size, robustness, availability, and so on. In order to deliver solutions covering all different needs, we are able to design, develop and deploy our technologies on a wide range of hardware solutions and leveraging most advanced frameworks and software tools for AI and Machine Learning.
α-MaLe™ technology can be adapted and customized to be run on very different devices, from embedded platforms for low-power & low-weight applications, to easily scalable cloud infrastructures, as well as more standard desktop solutions. All of them are developed using the very same set of best-in-class tools, guaranteeing portability, robustness and reliability.
Our experience allows us to support our customers in performing optimal trade-offs when identifying best solutions in terms of both hardware and software, reducing time-to-market and providing production ready applications adopting state of the art AI and Machine Learning.
Natural Language Processing
With an exponential growth of digital content, from social media to blogs to messaging apps, the ability to automatically extract and use information from it, has become a key goal.
α-MaLe™ implements a wide set of functionalities capable of carrying out complex processing on natural language content, with customizable features to accommodate specific requirements.
Semantic Analysis and Information Extraction
With modern machine learning tools it is possible to analyze a given content and recreate/abstract its conceptual structure, as well as extract and understand information contained in it.
In a world where vocal interaction with machines is becoming more and more common, speech recognition capabilities are of primary importance. State of the art modules available nowadays allow to integrate such features flawlessly.
One of the most attractive aspects of machine learning application is the capability of automating tasks and processes. An incredible amount of everyday activity, currently carried out by humans, can be automated and managed by smart agents. Once appropriately trained, these machines are able to perform their tasks more efficiently and more precisely than humans, increasing both product/process quality and profitability.
α-MaLe™ provides a large collection of software modules to automate many different activities. Its adoption allows to optimize processes boosting efficiency and returns.
One typical application is automatic integration of different services, where the autonomous agent learns how to link triggering events to generate new actions to be executed by other players.
In the Big Data Era, making sense out of enormous amount of information is critical. Machine Learning is the perfect tool to extract descriptive information from data, focusing on most important aspects.