Source: Deep Learning on Medium
Affirmative Learning: A new Machine Learning approach toward the AI
‘’ We must be careful not to believe things simply because we want them to be true. No one can fool you as easily as you can fool yourself.’’ Richard Feynman.
A decade ago, teaching a machine to recognize an object or an entity from an image involves a complete “features design” process. This designate a labor work where, in the case of recognizing a human body for example, you should define all the distinctive features that allows to recognize visually an entity human: The body shape, parts, etc. Teaching machines to recognize patterns was a painful process following the complexity of the real world where ML should perform (features complexity, unlimited number of exceptions, etc).
Then, Deep Learning happen. At the very high level, Deep learning was a major leap in the machine learning and in computer science in general. In fact, this family of approaches tends to automatically recognize features by learning their distribution from a large amount of data. For example, if I would like to teach my machine the ability to recognize cats from any RGB picture, I should expose my deep learning algorithm to a large number of various observations (different pictures of cats) in order to constitute a good internal representation of the entity “cat”.
Beyond the Deep Learning
After training a machine to recognize a pattern, data scientists and engineers generally follows the following popular and non exclusive avenues to improve their system accuracy:
(a) Improving the Algorithm itself: Improving the learning algorithm features representation and learning capabilities: New algorithms, Hyper parameter fine tuning, etc.
(b) Data enrichment: Using more and various data to train the learning algorithms.
We can notice that all of two approaches are improvement setting that are prior to the execution of the Machine Learning system. In fact, the interventions aims to make the AI system ready for a posterior execution in an uncertain and uncontrollable environment. This will suppose that during the execution of the AI system in a real environment, the machine learning system is not set for a short term objective self-adjustment. This is where we identified a room for improvement.
Learning algorithms are
freeze into wild exercises
We build and train algorithms to recognize pattens mainly in the wild. This phase, most known as the inference phase, consists mainly into mapping real world observations to previously learned representations in order to recognize them.
To recognize a cat in a camera footage, a trained algorithm will process a specific frame in order scout for learned features that could activate an internal representation of “cat entity”. It’s evident that even a well trained algorithm can still fail to recognize an object simply because there’s an infinite number of exceptional situations that are new from an algorithm perspective: long distance, very particular angle, unforeseen lighting, etc. The real environment is full of uncertainties that make AI naturally fails in uncontrolled conditions, like our world!
How biologically intelligent agents (like humans) deal with that?
Affirmative Learning: The intuition
To introduce you the affirmative learning concept, think about the following situation: there is a tiny round metal object sitting on the far extremity of table. It looks like a penny cent or maybe… a metal coat button. What you naturally do to get an affirmation about this? You may came closer, or move your head, you may stand up from your chair to get a better perspective, then you collect a better observation that could have a better probability to fulfill an affirmative answer.
Driven by their affirmation desire, any biologically intelligent agent (humans, some of the mammals, etc) instinctively setup their motricity capabilities by executing a sequence of actions that aims to acquire “a better observation” of the object.
If we refer to the machine vision area, this could means a controllable acquisition device (a camera for example) that will actively scout for the missing features when it fail to recognize an object with a specific threshold of certainty. The affirmative learning process will be autonomously activated when a predictor algorithm P (like Neural Net based classifier for objects) is able to recognize an object with a low confidence rate.
Affirmative Learning: RL driven data acquisition
The affirmative learning intuition was inspired (roughly) from the way that biological agents deal with the situation of incertitude. When we desire more certitude, we proceed into an extended data collection.
Following this sense, our intuition is to take advantage of the capability of the data acquisition device to acquire data, and this driven by a specific (missing) features need. To achieve that, we relay on a reinforcement learning algorithm that we designated as the Assistant A. This reinforcement learning agent will control the data acquisition device in order to collect a better observation that allows the Machine Learning algorithm Predictor P to recognize an entity (or any pattern in general) with a higher confidence.
Affirmative Learning: Use case and market insight
Let’s say that we’re controlling a PTZ camera: Pan–Tilt–Optical Zoom. We can map the finite action space (A) of the reinforcement learning based Assistant A with the possible discrete controls action that the camera dispose of. This could include more than motor actions like: Contrast, Light, etc. The assistant A will be coupled to a Predictor P in order to proceed to a targeted data collection loop on demand following the user set value of the certainty threshold. Trained collect the desired features that the Predictor P is missing, the Assistant A is trained to collect the features in a specific environment with a minimum number of actions.
It’s important to notice that affirmative learning is a framework, that designate the interaction mechanics between the Assistant A and the Predictor P. This means that the framework is agnostic to the specific algorithms of the Predictor or the Assistant. In fact, the nature of the predictor and the assistant could vary following the exercise objectives and the environment of execution.
To have a better sense of the market opportunity around this technology, let’s recall the following facts:
- Automation and manufacturing advancements resulted with affordable electronics. Since the price compression, they invaded every consumer part of the market.
- Connectivity democratization pushed the electronic brands to shift from to disconnected things to the Internet of Things.
- Connected devices opened a tremendous market of AI use cases
We see affirmative vision as raising in the intersection of the mentioned realities. In fact, we see a future where every object could be connected to channel informations and orders. Affirmative vision will be an additional layer of intelligence that will offer to those objects an autonomy regarding the continuous improvement regarding the tasks we will assign to it. A connected camera that have the responsibility to identify cars plates, will have an autonomous way to retrieve from failure by collecting the best possible observation, then by mastering the environment where it exists.