Universal Artificial Intelligence Tasks
Abstract- In view of all the ever-increasing data clouds, keeping a count of training points is no more essential, what one needs is a way in which one Machine Learning (ML) model can talk to other ML models, to start with. Sequentially, one should aim to merge various AI problems, this may be referred to as “model interoperability”. This article aims to explain the view to merge smaller ML tasks, to start with classification tasks with other ML/classification tasks, with the aim to make “many model – classify all”. The many model – classify all aims to work on interconnecting all the classification tasks at hands, for any kind of input and to recognize what the input is all about, for instance in crime detection. In this example, given a video a full story needs to be analyzed and summarized, this require, audio, image, Wikipedia, web and even medical knowledge bases to be present in the Universal Classifier. This is the countable classification task, as all we know is the ever-increasing number of models is enumerable and hence countable, even though the input space is itself infinite, since each individual input feature takes values from R infinite. The countable number of objects in real world, many, say kind of leaves, being determined by a p-class model, another, say kind of textures on a new Planet, depicted by m-class problem, so on.. Even the number of models would be increasing as time passes by, and these can be accommodated in the architecture of many models – classify all problem, since we are building this architecture with an aim that the number of models shall always be countable and classification shall be countable, as of now. As of now, is used as one day we shall have a fuzzy classification possible with the countable classifiers giving in a real valued membership value of infinite cardinality and fuzzy quantities, leading to number of target classes to be more than just countable, but an infinite classification task. The emphasis is to lay a foundation for models to interact with the proposed Universal Classifier in a collaborative mode.
Keywords: Infinite classification task, classification, Machine Learning, Artificial Intelligence, Future, Architecture, Universal Classifier
With the increase in number of problems we are solving in Artificial Intelligence (AI) day by day, the number of classification tasks are increasing at a phenomenal rate. For each image recognition task there is a neural model, for every supervised natural language processing task there is a model, same is true for MRI data, speech synthesis, medical data, financial or economic data, and in every field where AI is used to classify. The rate at which the data is increasing is a steady rate now. The rate of increase of this data may even increase with time or may decrease with the presence of some certain economic conditions, hence itself a relative concept. This is seeing it as every new year, many more classification tasks are studied, some are re-worked for improvement and while other tasks are novel, depending on the new challenges faced. The new challenges can be new problems altogether at hand such as the response of user for a new news, a prediction for winning of a new cricket match, may be the classification of new asteroid striking another heavenly body in Universe. These are just some of the many problems which are growing in number day by day. Can you count it? And how many classification classes can they fit in? What if we need to do the Fuzzy Classification with a real valued membership measure assigned to the classification classes, then what is the image of this classification mapping? It may be at times in a broad way, shall be a connected metric space, which is uncountably infinite, since non-empty.
Not only the problems are growing, the models are growing as well, day by day. There may be for sure models that are becoming obsolete as AI algorithms are maturing. The data itself is getting updated constantly, when new features are added, however this does not make original data obsolete, as compared to models. The data hence can be considered as to be increasing in volume each hour each day. For example, the weather data is being appended in the collection of weather data day by day, location by location, year by year. The data per department – say weather department only, is increasing in huge numbers. The count of this can be computed depending on if data is computed minute by minute or hour by hour. A finite value can be assigned to cardinality of such a data collection. Given such huge historical data only thing required is to comprehend it with help of AI tools.
There are problems which occur in real world data such as a merged field of Image Processing and Natural Language Processing, this involves the application of problem in field of video processing, video analysis and video summarization, to mention a few. These video images can be in space, earth, moon or even in movies – real or with other characters. These are problems that are growing at a tremendous rate day by day. Each frame of image with the speech in it is itself a grid and collection of pixels and sounds captured in that frame. These data are fed in all kinds of models in many kinds of classification problems. The classification problems can be two class classification problems or multi class classification problems. But consider the plethora of data collected in all these problems. When one counts in the data points for a cat image recognition classification problem, certain training data are provided. The model learns, the same way a bell image recognition problem, certain training data are provided, by manual collection and tagging. When an image with “cat with bell in neck” is obtained what does the model predict it the image as? When this image with “cat with bell” is fed to cat classification problem at times it shall predict it as cat and in minor cases it may predict it as a bell. Now, when the same image is sent to the model trained to detect bell, it shall classify it as bell. However, multi classification tasks shall predict this image as both – “bell” and “cat”. This problem does not stop here, as sooner we may like to know what kind of cat it was – a spotted cat or a furry cat, a think cat or a kitten, to even the kind of bell and color of the bell and where exactly the bell is in the frame so collected. We have an ever-increasing task set at hand, given the increasing number of objects we want to recognize with help of image processing and so many expressions we want to learn with help of natural speech analysis. Yes, with speech recognition, we want to know not only what is expressed but with what sentiment value (a real number) of expression it is expressed, this adds yet more dimensions to the problem classification features as well as the target to be projected out. This data is increasing at a rate which can’t be counted to exact number as lest it should not be same number the next minute, this is increasing at a steady rate.
In face of all these data clouds available, keeping a count of training points is no more essential, what one needs is a way in which one Machine Learning (ML) model can talk to other ML model, to start with. Sequentially, once should aim to merge various AI problems, this may be referred to as “model interoperability”. This aims to merge smaller ML tasks, to start with classification tasks with other ML/classification tasks, with the aim to make “many model – classify all”, a Universal Classifier. The many model- classify all aims to work on interconnecting all the classification tasks at hands, for any kind of input and to recognize what the input is all about. This is the countable classification task, as all we know is the number of classes are enumerable with Natural Numbers and hence countable. Even the number of models would be increasing as time passes by, and these can be accommodated in the architecture of many models – classify all problem, since we are building this architecture with an aim that the number of models shall always be countable. All this to view the problem at hand viz. to build the Universal Classifier, in which neural and other models can be plugged in and even plugged out to be replaced by a newer better version. In many cases, the training data can be infinitely uncountable as well, for example, the analog signals and speech recognition tasks, when accumulated over time, with no bounds on time. This is not the current problem at hand since at present we can enumerate it to countable problem, given the current measurement apparatus is such, was it predicted even the infinite training data accumulating in our data warehouses, so how can one rule out the other, though too early for it as of now.
Why infinite classification problem and no longer a finite problem ? The reason as we try to fuse models which are learned over a finite training data with say n1 features the problem is in space of RN1 as the inputs are real-valued and hence the evaluation can vary over RNI. Then there is second model learned over a completely different number of features say n2 features, this is a problem in RN2 space and so on…… If we have p models then with np features then this is a problem in RNP space. The accumulated model works on aggregation of all these models and is in space of RN1 * RN2…*RNP. This is the Cartesian product. Given we shall be building an architecture to make models so that they can be fused on and an input can be fed in to all models, some may recognize it while others would learn to return a null statement as output, indicating that the input is not recognized by the models. To start with the we shall be making the architecture of the classify-all problem form a finite number of models, but as time progresses, more and more models shall be included in the architecture. The architecture must be built in a way such that – any new model can be plugged into it, which is to say, any new neural model can be put in the Universal Classifier architecture, fed in the same input and learn to return its output to the architecture or return null. The infinite problem arises as – we would no longer be counting on how many models are plugged into system. The architecture should be built in such a way, that more and more models can participate in this. Now, this rises the complexity issue, well we have Quantum Computers now! And the future even more highly processing machines to run in parallel, most models shall return null, on recognition of input only, hence won’t be processing the model as well, thus exiting on the first go itself, saving lot of computational powers. The rest shall process and recognize and send its outputs to the global output collector. Fuzzy membership values makes it even more difficult to enumerate finitely.
The Universal Classifier works initially based on majority votes for crisp Universal Classification while works on Fuzzy Aggregation for Fuzzy Universal Classification Tasks. This shall be the basis to compute the summary of any real world data which the Universal Classifier shall comprehend. For example, the aim of say crime detection, shall take input as a video, and it shall take all the models possible in world as plugins, and shall convert to natural language based on majority-votes of all classifiers which distinctly summarize the scenes in the video for automatic crime detection. This is just a start for the Universal Machine Learning Tasks.