Imagination based AIs — Part -2
Imagination Augmented Learning Agents(I2A)
Generally known as Imagination Augmented learning agents or I2A, Imagination based AIs are neuroscience motivated and their algorithms are based on human brain’s ability to solve problems based on previous experiences. Unlike other machine learning algorithms, these algorithms do not use available data and predictive analytics, instead use the optogenetics[UC1] and brain’s hippocampus map to achieve conceptual learning based on imagination and reasoning based planning to solve the real life problems. This type of AIs can create elaborate plan from the existing and previous experiences in sequence and they can be efficiently linked together to make the best ‘step by step’ plan to execute a task. The plans made, involve imagination and reasoning on the available information gathered from these experiences. The algorithm is structured like a tree [UC2] linked from one branch to another and the decision is made to follow a branch based on the association between the input, reward, and the task/goal.
The idea of it all is extremely simple.
· The AI system takes an input from the environment or the available data.
· The machine learning algorithm also known as an ‘agent’ is fed with this input. It uses this input and its creative skills and arrives at a state of action.
· Then the action is implemented, and it is represented as the output of the AI system.
· Every action is a step taken towards executing the task or reaching the final goal.
· Once the action is performed, then a feedback loop is used to check the status of the step taken towards the goal.
· If the step taken is a wise one, it yields a reward and does not get penalized.
· Then onwards, the algorithms proceed to take the second, third, fourth …. steps, towards the goal.
· But in case, if the first step taken is a wrong one, it gets penalized. Now the algorithms change the output accordingly and then the feedback loop checks the first step once again.
Imaginative AIs also uses deep reinforcement learning. Reinforcement learning aims to obtain complex goals, by maximizing their focus on a specific dimension or direction, obtained over a number of iterations. The input is defined by the initial state and the output is the next state, which earns a reward. The algorithm trains model mainly based on input state and the feedback received from the previous state. In every state or step, the learning agent learns to adapt, till it reaches the goal with maximum reward. The agent is penalized for making a move in the wrong direction and rewarded for a move made in the right direction. The right and the wrong directions are learnt over a period of time through many iterations, by making mistakes and the algorithm which makes the least mistakes and achieves maximum rewards are the best reinforcement learning algorithms.
This is similar to how children learn to reach a candy jar, which is placed on the top kitchen shelf. It may not always happen in the first instance. It might take a few turns for the child to learn to walk to the kitchen shelf without their parents noticing the action. And it may take a few more instances to obtain a small step-stool or chair to reach the top shelf, before they find the candy jar. The process might take a few days to master or a few hours to master, depending how smart and intelligent the child is and how quickly they learn and adapt to the environment, they are in. Though the goal is the candy jar, the rewards can be two or three items, such as
· To reach the kitchen taking minimum number of steps from the living room.
· To find the small tool or chair, without making a sound or without dragging the attention of the pet dog sleeping in the kitchen.
The rewards may vary based on the task, input and action taken to change the state from input to output. These, rewards can be used as a tool to reach the final goal, of grabbing the candy jar. Also, the child may be grounded or punished by their parents if they do something which make a sound or drag their attention, which is equivalent to a penalty.
Types of Machine Learning Algorithms
The main difference between traditional algorithms and the machine learning algorithms which are used by Artificial intelligence software is its predictive nature, while the similarity being both algorithms involve the processing power of the computing systems. The standard algorithms are programmed to follow some pre- defined rules in order to make decisions and complete the task using the given inputs. Nonetheless, the ML algorithms are trained to use the available data to predict and analyze the given input. There are three types of machine learning algorithms. The main difference between the three is the input data and the training concept or learning style and the training data.
1. Supervised algorithms use a group of images, texts or moving video to capture the data and classify the inputs to complete the task based on those images, texts, and video.
2. Unsupervised Algorithms uses the various data and group them and classify them based on the similarity features of the images, texts or video using feature extraction, also known as dimensionality reduction, in order to complete the task.
3. Reinforcement algorithms uses the data from the environment to learn, adapt, make decisions, and receive rewards in completing the task.
While supervised learning algorithms trains the model using the classification of object itself, reinforcement learning, does not have an answer object. The task is to figure out the answer. For example, an image of cat is given as the key or the answer and the supervised algorithm is expected to find all the cats among the given input images, by using similarities in the feature. Reinforcement learning is a nearer cousin to unsupervised learning than supervised learning. Nonetheless, in unsupervised learning the classification or group is identified by pattern recognition methods such as clustering, reinforcement learning uses reasoning and imagination to perform the task.
Whenever I talk to my clients about using AI in IoT projects or IT security, they ask me to choose an algorithm for their specific needs. I want to mention a simple concept here about choosing ML algorithms for different requirements. Every algorithm whether it is regression, classification, clustering or Q- table, possess its own benefits in terms of learning style and shortcomings in terms of inductive bias[UC3] . So don’t be afraid to explore a few before deciding on the one. That’s the core idea of ML algorithms. There are no hard and fast predefined rules unlike the traditional processing algorithms, because the ML prototyping, is based on guesswork and intuition. So, my usual advice to them is always, ’follow your gut’. The best algorithm for a specific type of data can be found only, through analyzing the results spat out by three or four similar algorithms. Data is the key to all ML algorithm training. So, focus on your data and choose the best. There are plenty of cheat sheets available on the web for various ML algorithms. So, download a few cheat sheets and decide on two or three closest algorithms and run your data with them and decide on the best. Also don’t forget to use one of those AI platforms which eases the work of developing , training and testing the model.
Sentience in AI systems
Imaginative AIs are also equipped with sentience, the ability of an AI to perceive, reason, learn, adapt, and feel human pain and sufferings. Sentience can be divided into 3 major categories.
- Cognition: The ability of AI to understand and learn, reason, and adapt from its environment using its robotic sensors. Cognition contributes to the decision making ability of a sentient AI. The researcher in this field of AI sentience have already achieved 40% success. The focus of this series is about the cognitive abilities of Imagination based AIs. Hence, I will be discussing this topic throughout this series, in detail.
2. Consciousness: Human consciousness is simply based on our experiences. The longer we live, the more aware we are about our surroundings, our people, our family, and our friends. This refers to our feelings, awareness, memories, and environments. As per Ned Block, the famous author and American Philosopher, human consciousness can be of two types. Phenomenal consciousness is the awareness, while access consciousness refers to the actions performed based on the awareness gained. For example, when we see a stray dog or cat abandoned on the road, we decide to either take it home with us or to a pet shelter. The phenomenal consciousness allows us to feel the pain and suffering we visualize, while the access consciousness makes us perform the sympathetic act of adopting the stray or finding it a place to stay. As human beings, we take consciousness as granted, even if we do not have any type of medical science education or a scan or proof for our consciousness. Consciousness, in humans, is believed to be awarded by our five senses (sight, sound, touch, smell and taste). If all the five senses are awake, then we are deemed conscious. But it’s not that simple and that’s the reason it’s referred by biological and engineering researchers as the hard problem, because the human sixth sense called intellect or reasoning plays extremely important role in consciousness.
But again, it can be argued, about human pain and suffering being completely unrelated to the intellectual ability of a person. To cite an example, I have watched ‘David Attenborough’s series on Planet Earth’, where elephants in a herd moan over the death of a herd-mate, standing in a circle around the dead animal. They understand emotions, such as fear, pain, hunger, and thirst, without the sixth sense of intellectuality. Similarly, it is quite common for a dumb person to be highly emotional, and the same emotion can be felt by a person with high IQ. Thus intellect, learning and reasoning has nothing to do with consciousness.
If you ask Siri, ”Hey Siri, how are you today?” Siri, immediately answers, “I am happy to be here”. But it is not a true answer, because Siri does not understand the difference between happiness or sorrow, but it is trained by its programmers to fake an emotion and return the answer as an exchange of pleasantries, to start a conversation. But as I mentioned in my previous article, ‘Siri’, belongs to the classification of AI called narrow AIs, and their learning and adapting is strictly restricted to a given set of data. Thus, Narrow AIs like Siri, are not Sentient AI, as they are not conscious, but they are made to simulate consciousness through understanding human thoughts.
Again, it is not that simple to draw a one to one comparison to human or animal consciousness to machine consciousness, for a simple reason, machines are equipped with more stable memory than human beings or animals. Machines can store and access enormous amount of data, without understanding the impacts of the severity of the data collected and stored in its memory.
Thus, machine consciousness is defined by different researchers differently. Some researchers like David Levy got so fed up with finding the right definitions to explain machine consciousness has said,
“let us simply use the word and get on with it.”
If you’re interested to learn more about machine consciousness, click the link below.
Artificial Intelligence: Does Consciousness Matter?
Artificial Intelligence: Does Consciousness Matter?Consciousness plays an important role in debates around the…
This brings us to the end of Part -2 of this series. Watch the same space for more info on this subject. If you’re not already following me on medium.com, do it now.
[UC1]Optogenetics is the recent field of study, (started in 2004) which defines vision based perception. This field uses optics or light control, to trigger an activity in human brain cell and this controls the decision making process of human brain. In simple words, ‘it is what the eye sees and how the brain interprets it’
[UC2]Decision trees are one of the best supervised machine learning algorithms which uses the data collected to create patterns. They are trained to classify an information into one group or the other based on branch of the tree the algorithm ends up with. They use two variables X and Y to take a decision. In general, the X variable is the parent of the Y variable or the dependent variable.
[UC3]Inductive bias in ML algorithms happens due to some assumptions made while the inductive learning is enforced. I am planning to write an article on Inductive bias and hypothesis space in ML algorithms shortly. Follow me on medium.com for more info